Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
Dis cus si on Paper No. 08-128
FDI and TaxationA Meta-Study
Lars P. Feld and Jost H. Heckemeyer
Dis cus si on Paper No. 08-128
FDI and TaxationA Meta-Study
Lars P. Feld and Jost H. Heckemeyer
Die Dis cus si on Pape rs die nen einer mög lichst schnel len Ver brei tung von neue ren For schungs arbei ten des ZEW. Die Bei trä ge lie gen in allei ni ger Ver ant wor tung
der Auto ren und stel len nicht not wen di ger wei se die Mei nung des ZEW dar.
Dis cus si on Papers are inten ded to make results of ZEW research prompt ly avai la ble to other eco no mists in order to encou ra ge dis cus si on and sug gesti ons for revi si ons. The aut hors are sole ly
respon si ble for the con tents which do not neces sa ri ly repre sent the opi ni on of the ZEW.
Download this ZEW Discussion Paper from our ftp server:
ftp://ftp.zew.de/pub/zew-docs/dp/dp08128.pdf
Non-technical summary
Despite the continuing political interest in the usefulness of tax competition and tax coordina-
tion as well as the wealth of theoretical analyses, it still remains open whether or when tax
competition is harmful. Moreover, the influence of tax differentials on multinationals’ deci-
sions is still insufficiently analyzed. Thus, economists have increasingly resorted to empirical
analysis in order to gain insights on the elasticity of foreign direct investments (FDI) with
respect to company taxation. As a result, the empirical literature on taxation and international
capital flows has grown to a similar abundance during the last 25 years as the respective theo-
retical literature. Its heterogeneity leads to a rising need for concise reviews on the existing
empirical evidence. In this paper we extend former meta-analyses on FDI and taxation in
three ways. First, we add the most recent publications unconsidered in meta-analyses up-to-
date. Second, we apply a different methodology by using a broad set of meta-regression esti-
mators and explicitly discuss which one is most suitable for application to our meta-data.
Third, we address some important issues in research on FDI and taxation to the clarification
of which meta-analysis can make valuable contributions. These issues are mainly: The influ-
ence of variables which might moderate effects of tax differentials (e.g. public spending), the
implications of using aggregate FDI data as opposed to firm-level information on measured
tax effects, the implications of bilateral effective tax rates, and the possible presence of publi-
cation bias in primary research.
According to our meta-analysis, a pooled effect based on the median result taken from each
primary study could be obtained that amounts to a semi-elasticity for company taxes on FDI
(percentaged reaction of FDI to one percentage point change in the tax burden) of 1.68 in ab-
solute terms.
We do not find overwhelming support for publication bias in the meta-regressions. The use of
aggregate data leads to higher estimated semi-elasticities, but significantly reduces t-values
and thus implies less precise estimates. Regarding the choice of tax rates, unilateral effective
average tax rates do not lead to significantly higher tax effect size estimates or significances
as compared to the statutory rate. Bilateral tax rates best capture tax incentives and yield both
significantly higher effect size estimates as well as higher significances. Regarding the control
variables, it is most interesting that primary estimates are not significantly affected by the
inclusion of public spending. This might be due to the fact that only crude measures of public
inputs are used in many studies.
Zusammenfassung
Trotz des hohen politischen Interesses an den Themen des Steuerwettbewerbes bzw. der Steu-
erharmonisierung und der daraus resultierenden Vielzahl theoretischer Analysen zum Thema
bleibt die Frage nach deren ökonomischen Implikationen letztlich noch unbeantwortet. Insbe-
sondere der Einfluss des internationalen Steuergefälles auf die Investitionsentscheidungen
multinationaler Unternehmen ist deshalb zunehmend ein Objekt der empirischen Analyse
geworden.
Die empirische Literatur zum Zusammenhang zwischen Besteuerung und ausländischen Di-
rektinvestitionen (ADI) der vergangenen 25 Jahre ist indes von einer ähnlichen Diversität
geprägt wie die entsprechenden theoretischen Analysen. Diese Heterogenität der Ergebnisse
führt zu einem steigenden Bedarf an präzisen, methodisch fundierten und vor allem quantita-
tiven Übersichtsstudien (sog. Meta-Analysen) über die bestehende empirische Evidenz. In
dieser Arbeit werden vorhergehende Meta-Studien zum Thema Besteuerung und ausländische
Direktinvestitionen auf dreifache Weise erweitert. Zunächst wird die Datengrundlage der
quantitativen Meta-Analyse um die aktuellsten Studien bereichert. Ein ausführlicher methodi-
scher Teil der Analyse wird anschließend unterschiedliche Schätzverfahren zur Durchführung
von Meta-Regressionen in einem integrierten ökonometrischen Rahmen darstellen und zudem
eine Strategie zur Auswahl des zur Analyse der vorliegenden Meta-Daten präferierten Verfah-
rens entwickeln. In einem dritten Schritt werden anschließend wesentliche Aspekte unter-
sucht, deren Diskussion in der aktuellen empirischen Literatur von Bedeutung ist. Die Metho-
dik der Meta-Analyse ist geeignet, für zahlreiche dieser Aspekte einen klärenden Beitrag zu
generieren. Zu den im Detail untersuchten Fragestellungen gehört der potenziell mildernde
Einfluss der Ausgabenseite des öffentlichen Budgets auf den (abschreckenden) Steuereffekt.
Zudem werden die Implikationen der Verwendungen aggregierter Daten (im Gegensatz zu
Informationen auf Firmenebene), die Implikationen der Verwendung bilateraler Steuerbelas-
tungsmaße (im Gegensatz zur Verwendung unilateraler Steuerbelastungsmaße ohne Berück-
sichtigung bilateraler steuerlicher Regelungen) sowie die mögliche Existenz einer publikati-
onsbedingten Verzerrung der in der Primärliteratur gemessenen Effektgrößen (sog.
publication bias) untersucht.
Im Ergebnis ermittelt die vorliegende Arbeit u.a. eine präzisionsgewichtete durchschnittliche
Semi-Elastizität (prozentuale Reaktion der ADI auf eine Variation der Steuerbelastung um
einen Prozentpunkt) von absolut 1.68, basierend auf den Median-Resultaten aus jeder der be-
rücksichtigten Studien. Keine eindeutige Evidenz findet sich für das Vorliegen von
publication bias. Indes scheint die Verwendung aggregierter Daten die gemessenen adversen
Effekte der Besteuerung auf ADI zu verstärken, die Präzision der Schätzung jedoch zugleich
zu verringern. Bilaterale Steuerbelastungsmaße führen zu (absolut) höheren primär gemesse-
nen Steuereffekten als ihr unilaterales Pendant. Sie scheinen somit eher geeignet, die von der
Besteuerung ausgehenden Anreize auf transnationale Investitionsentscheidungen zu erfassen.
Interessanterweise konnte kein mildernder Effekt von Größen auf der Ausgabenseite des öf-
fentlichen Budgets auf die abschreckende Wirkung der Besteuerung nachgewiesen werden.
Dies könnte als Hinweis auf die bisherige Verwendung unzureichender Maße für die Bereit-
stellung der aus Unternehmenssicht relevanten öffentlichen Güter gewertet werden.
FDI and Taxation
A Meta-Study
Lars P. Feld*, Jost H. Heckemeyer**
December 2008
Abstract:
Despite the continuing political interest in the usefulness of tax competition and tax coordina-tion as well as the wealth of theoretical analyses, it still remains open whether or when tax competition is harmful. Moreover, the influence of tax differentials on multinationals’ deci-sions is still insufficiently analyzed. Thus, economists have increasingly resorted to empirical analysis in order to gain insights on the elasticity of FDI with respect to company taxation. As a result, the empirical literature on taxation and international capital flows has grown to a sim-ilar abundance during the last 25 years as the respective theoretical literature. Its heterogenei-ty leads to a rising need for concise reviews on the existing empirical evidence. In this paper we extend former meta-analyses on FDI and taxation in three ways. First, we add the most recent publications unconsidered in meta-analyses up-to-date. Second, we apply a different methodology by using a broad set of meta-regression estimators and explicitly discuss which one is most suitable for application to our meta-data. Third, we address some important issues in research on FDI and taxation to the clarification of which meta-analysis can make valuable contributions. These issues are mainly: The influence of variables which might moderate ef-fects of tax differentials (e.g. public spending), the implications of using aggregate FDI data as opposed to firm-level information on measured tax effects, the implications of bilateral effective tax rates, and the possible presence of publication bias in primary research.
JEL-Classification: H25, H73, F23, F21.
Keywords: Corporate Income Taxation, Foreign Direct Investment, Meta Analysis.
We are extremely grateful to Ruud de Mooij for generously providing his data set to us. We also would like to thank Laura Kaeding, Ivan Mitrovic, Benjamin Protte and Frank Schneider for very valuable research assistance.
The authors also would like to thank Martin Paldam, Doina Radulescu, Gonzague Vannoo-renberghe and other participants at the IIPF Conference in Maastricht, the MAER Workshop on Meta-Analysis in Nancy and the joint Ph.D. workshop in Mannheim for helpful comments. * Prof. Dr. Lars P. Feld holds the Chair of Public Economics at Heidelberg University and is research associate at the Centre for European Economic Research (ZEW). ([email protected]) ** Jost H. Heckemeyer is researcher at the Centre for European Economic Research (ZEW). ([email protected])
- 1 -
1. Introduction The political interest in tax competition, coordination and harmonization has remained high in
Europe for more than 45 years (see the Neumark-Report published in 1963).1 After successful
coordination and harmonization in indirect taxation, company taxes attract attention in en-
suing debates. Most recently, the EU considers a harmonization of corporate taxation in form
of a common consolidated corporate tax base (CCCTB) combined with formula apportion-
ment (Martens-Weiner 2006, Schön, Schreiber and Spengel 2008). Not surprisingly, the
scientific interest in international tax competition has considerably increased since harmoniza-
tion efforts in Europe intensified, but also due to rising capital mobility in the last thirty years.
Theoretical studies in public economics provide the conditions for tax competition being
harmful and tax coordination being useful. These conditions are however numerous and hete-
rogeneous (see the surveys by Wilson 1999, Feld 2000, Wilson and Wildasin 2004, Fuest,
Huber and Mintz 2005). Without referring to empirical evidence, no unambiguous predictions
could be obtained. As a result, the empirical literature on tax competition has heavily grown
in recent years sadly leaving us with a similarly diverse picture as the theoretical analyses (see
the surveys by Hines 1997, 1999, Devereux 2006). In contrast to theoretical analyses, econo-
mists could however adopt meta-analytical methods from the life sciences to gain clearer in-
sights as to the effects of international taxation instead of the usual narrative surveys on em-
pirical results (Jarrell and Stanley 1989, 2005, Stanley 2001). De Mooij and Ederveen (2003,
2005, 2006) pioneer this approach in the economics of international taxation.
As the role of multinational firms and problems of financial arbitrage are increasingly focused
in the theoretical analyses on the impact of international tax differentials on foreign direct
investment (FDI) (Fuest, Huber and Mintz 2005), the multinationals’ options for location
choice are the starting point for empirical analyses. Dunning (1981) identifies three advantag-
es, called O-L-I, that have to be present for a multinational to invest in foreign host countries.
Besides locational advantages (L) offered by the host country, to which taxation belongs, the
multinational should be able to realize firm-specific advantages as compared to domestic en-
terprises in order to successfully compete with them despite of costs of setting up activities in
a foreign country. The ownership advantages (O) must therefore be complemented by an in-
ternalization advantage (I) over, e.g., licensing of foreign production.
1. Debates about the effects of tax competition certainly date back much earlier, at least if the situation within
countries is additionally considered. Spoerer (2002), e.g., provides historical accounts of tax reform discus-sions in Zurich in 1883 pointing to the “flights of the capitalists” induced by the enormous tax differential to the canton of Basel (see also Georg von Schanz 1890). Another historical study extending the analysis of in-terjurisdictional competition into pre-modern Germany between 1000 and 1800 stems from Volckart (2002).
- 2 -
This broad perspective on FDI allows for a consideration of other factors influencing multina-
tionals’ investment decisions. In particular, studies in international economics distinguish
horizontal expansions from vertical investment decisions and analyze both issues separately.
Multinationals that engage in horizontal investments want to establish production abroad in
order to serve local markets. In contrast, vertical investments serve to geographically allocate
a company’s production chain with the aim of exploiting inter-locational differences in rela-
tive factor endowments. Based on these insights, Horstmann and Markusen (1992) develop an
instructive theory on horizontal FDI. They show that a multinational will locate production in
direct proximity to a foreign market if the corresponding reduction of distance-related costs is
advantageous as compared to the reduced potential of exploiting economies of scale at the
plant-level. The more important economies of scale are, the less likely local production will
become (at least in small foreign countries). As to vertical FDI, Helpman (1984, 1985) ex-
plains the splitting of the value chain as being driven by the multinational’s aim to reduce
total production costs. As long as reductions in production costs, due to lower factor prices
abroad, exceed corresponding set-up costs, the multinational firm will engage in foreign (un-
skilled-labor intensive) production. Eby-Konan, Markusen, Venables and Zhang (1996) and
Markusen (1997, 2002) finally present a unified approach. Their so-called knowledge-capital
model explains both, horizontal and vertical, motivations for foreign investment. This model
has become the standard theoretical approach in FDI research.
This theoretical perspective on the multinationals’ investment decisions in a world of interna-
tional tax differences underlines the importance of additional factors influencing FDI and thus
of the ceteris paribus assumption in tax competition analyses. Some recent empirical studies
on FDI and taxation pursue the above-mentioned concepts in order to derive a theoretically
well-founded estimation strategy.2 However, many studies instead opt for a traditional gravity
set-up as successfully applied in the empirical trade analyses. The gravity approach explains
inter-country FDI patterns by a combination of mass variables (e.g. GDP, population) and
distance variables. This simple set-up is regularly augmented by presumably important policy
and locational factors. These are – in addition to a measure of taxation – indicators for politi-
cal risk, openness, infrastructure, labor costs, etc. Thus, they often employ sets of explanatory
variables whose possible importance is also justified by theoretical considerations.
In addition to the underlying theoretical concepts, empirical studies on FDI and taxation vary
widely with respect to specification, sample size and so on. One very important aspect in
2. Early empirical studies implicitly relied on q-style investment models (Hines 1999).
- 3 -
which they differ is the dependent variable. Since FDI in its broad definition does not only
comprise real investments but also financial flows due to mergers or acquisitions of already
existing capital, some authors prefer to use a dependent variable which better reflects real
economic activity. They therefore rely on investments in stocks of property, plant and equip-
ment (PPE) which are undertaken or held by foreign affiliates in a particular host country.
The central explanatory variable, i.e. the tax variable, also varies on the basis of different al-
ternative concepts. Besides the use of statutory tax rates at the country or regional level, many
empirical studies rely on average tax rates. In contrast to statutory tax rates, average tax rates
reflect the tax base to which statutory tax rates apply. They are computed on the basis of pre-
vious microeconomic or macroeconomic data and are therefore also known as backward-
looking tax measures. One special feature of backward-looking tax measures is that they cap-
ture the “true” tax burden on investments or companies. Since they are based on actual tax
payments they reflect the outcome of tax planning and discretionary tax provisions. However,
one of their major caveats also lies in this backward-orientation. As Devereux (2006) points
out, average tax measures might introduce an endogeneity bias into empirical analyses. Since
they are based on previous data, average tax rates might be influenced by recent investments
(e.g. through depreciation allowances). The direction of causation between the dependent FDI
variable and the average tax rate thus might be ambiguous.
From a theoretical point of view it is preferable to use tax measures which are forward-loo-
king and thus better reflect the situation investors are in when they decide. Forward-looking
tax measures are differentiated into effective marginal and effective average tax rates. Effec-
tive marginal tax rates measure the tax-induced relative wedge between pre- and post tax re-
turns on marginal investment. They are mainly calculated using neoclassical investment mod-
els developed by King and Fullerton (1984) or Devereux and Griffith (1998, 1999). Devereux
and Griffith (1999) also calculate effective average tax rates, which measure relative tax-indu-
ced reductions of the net present value of inframarginal, i.e. profitable, investments.3 These
measures reflect the major provisions of the considered tax system in a transparent and com-
pact way. They most importantly correspond to the forward-looking nature of investment de-
cisions which are taken on the basis of future taxes triggered by the respective projects.
The meta-analyses by De Mooij and Ederveen (2003, 2005, 2006) particularly study to what
extent the choice of tax rates and the differentiation in the type of investment influence the
3. Effective average tax rates might also be calculated by means of modeling a representative company’s fi-
nancial development over a certain period of time. See Spengel (1995) or Jacobs and Spengel (1996).
- 4 -
results of empirical studies on tax competition. They put less emphasis on the control va-
riables which are used in those studies. Most importantly, the impact of public spending on
multinationals’ investment decisions has received little attention there. Adopting the tradi-
tional Tieboutian (1956) perspective on inter-jurisdictional tax differences, the spending side
must however not be neglected. This has been recognized by theoretical studies (Keen and
Marchand 1997, Sinn 2003) and much emphasized by Bénassy-Quéré, Gobalraja and Trannoy
(2007) in their recent empirical analysis. Similarly, Baldwin and Krugman (2004) show that
agglomeration effects in geographic centers might allow governments to tax business more
heavily. In a recent unpublished paper, Brülhart et al. (2007) present evidence for newly
founded firms that agglomeration effects actually reduce their sensitivity to tax differentials.
In this paper, we provide a meta-analysis on the impact of company taxation on FDI by par-
ticularly asking to what extent this impact is moderated by spending policies or agglomeration
effects. In addition to the use of a broader data set by inclusion of most recent studies, a me-
thodological expansion is made by employing a broad set of meta-regression estimators. We
explicitly test which estimator is most suitable with respect to the underlying data. Meta-
regression analysis may be based on fixed or mixed effect estimators, again to be distin-
guished into pooled WLS and cluster econometric methods. As shown by Bijmolt and Pieters
(2001) using random effects cluster estimation techniques which take account of the pre-
sumed dependence problem resulting from multiple sampling in meta-analysis might be ad-
visable. We choose among different methodological options on the basis of appropriate tests.
The remainder of the paper is organized as follows: In Section 2, we provide a qualitative re-
view of the empirical literature followed by a description of the data set derived thereof in
Section 3 and the descriptive statistics of the studies in Section 4. A classical meta-analysis
follows in Section 5, while Section 6 contains the meta-regression results. Concluding re-
marks follow in Section 7.
2. A Qualitative Review of the Empirical Literature
Since the early 1980s, a rich empirical literature has evolved disentangling determinants of
FDI with a special focus on the effects of company taxation. Following Devereux (2006), the
empirical literature on the effects of taxes on FDI allocation can be classified along three di-
mensions: the type of data used (time series, cross-section, panel data), the level of aggrega-
tion, and the stage of transnational investment decision analyzed in the relevant studies. In a
strict sense the decision stage dimension only becomes relevant in studies based on micro
- 5 -
data; only these are able to separately analyze discrete and continuous investment choices.
However, in accordance with De Mooij and Ederveen (2003) we use a slightly more general
classification and distinguish studies which explicitly examine discrete location choices from
studies using some sort of continuous capital data as dependent variable.4 Consequently, each
study can be assigned to one of 18 possible categories (Figure 1). It should, however, be
noted that not all imaginable combinations of study dimensions have been realized in the past.
Figure 1: Dimensions of empirical literature on FDI and taxation
2.1 Time Series Analysis on Aggregate FDI
Pioneering studies in FDI research used aggregate time series data (see Table 1). They ex-
amined the correlation of relative after-tax rates of return on U.S. investments and the annual
volume of inward FDI flows.5 This early approach by Hartman (1984) has been picked up by
researchers, which used longer time series and varied the underlying data as well as the speci-
fications (Boskin and Gale 1987, Newlon 1987, Young 1988, Murthy 1989). However, mod-
4. Micro studies using such capital data indeed analyse continuous investment decisions while studies based on
aggregate capital data generally compound location and continuous investment decisions. 5. For surveys that cover this strand of the literature in more detail see Hines (1997, 1999), De Mooij and
Ederveen (2003) and Bloningen (2005).
Continuous capital data
Time series Cross-section Panel
Location choice
Aggregate
Aggregate data on affiliates
Wijeweera et al. (2007) Bellak et al. (2007) Stöwhase (2005 Gorter/Parikh (2003) Billington (1999) Swenson (1994)
Wolff (2007) Benassy-Quere et al. (2007) Hajkova et al. (2006) Benassy-Quere et al. (2005) Broekman/Vliet (2001) Devereux/Freeman (1995)
Cassou (1997) Slemrod (1990) Murthy (1989) Young (1988) Newlon (1987) Boskin/Gale (1987) Hartman (1984)
Grubert/Mutti (2004) Swenson (2001) Devereux/Griffith (1998)
Egger et al. (2008) Demekas et al. (2007) Bellak/Leibrecht (2007) Razin et al. (2005) Buettner (2002) Shang-Jin (1997) Jun (1994)
Altshuler et al. (2001) Grubert/Mutti (2000) Hines (1996) Hines/Rice (1994) Grubert/Mutti (1991)
Firm-level data
Overesch/Wamser (2008a) Papke (1991)
Overesch/Wamser (2008b) Stöwhase (2005b)
Buettner/Wamser (2008) Desai et al. (2004)
Overesch/Wamser (2008a) Becker et al. (2006b)
Altshuler et al. (2001) Grubert/Mutti (2000) Hines (1996) Hines/Rice (1994) Grubert/Mutti (1991)
Mutti/Grubert (2004)
Buettner/Wamser (2008) Swenson (2001) Bartik (1985)
Buettner/Ruf (2007) Devereux/Griffith (1998)
Devereux/Lockwood (2006)
Bobonis/Shatz (2007)
- 6 -
ifications did not lead to differing results as compared to Hartman (1984). The established
elasticities with respect to after-tax rates of return did not exceed 0.5 in the case of transfers of
funds and ranging at around 1.5 for retained earnings. Only when Slemrod (1990) fundamen-
tally reconsidered the research done so far and, inter alia, put forward new controls as well as
a forward-looking tax measure, doubts were cast on previous results. Retained earnings were
not affected by taxes while transfers of funds showed significant reactions. In addition, sepa-
rately analyzing bilateral U.S. inward FDI from seven countries that differ in their regimes of
double taxation relief – as done by Slemrod – did not yield clear-cut insights about the influ-
ence of home country taxation on the tax sensitivity of FDI flows. Following Slemrod’s work,
Cassou (1997) is the latest paper to perform single time series analysis on bilateral FDI. Based
on a sample of six investing countries, he also constructed a panel and found a significantly
negative impact of the corporate tax rate on inward FDI.
Table 1: Empirical studies using aggregate time series data Study Data Special Remarks Tax
Measure Method Main results
Hartman (1984) US aggregate FDI inflow , 1965-1979
Annual inward FDI regressed on relative after-tax rates of return, transfers vs. retained earnings
BL OLS Elasticities with respect to after-tax rates of return do not exceed 0.5 (transfers) and 1.5 (retained earnings).
Boskin and Gale (1987)
US aggregate FDI inflow, 1956 - 1984
Extending Hartman (1984) with alternative data on rate of return
BL OLS Results in accordance with Hartman (1984)
Newlon (1987) US aggregate FDI inflow, 1956-1984
Criticizing Hartman (1984), coping with spurious correlation
BL OLS Results in accordance with Hartman (1984)
Young (1988) US aggregate FDI inflow, 1953 - 1984
Extending Hartman (1984), with lagged investment
BL OLS Results in accordance with Hartman (1984)
Murthy (1989) US aggregate FDI inflow, 1953 - 1984
Re-estimating Young (1988), with adjustment for autocorrelation
BL ML Estimator New funds show elastic reactions with absolute elasticity greater 1. Re-tained earnings also are more elastic than Young suggests.
Slemrod (1990) Bilateral US inbound FDI from 7 countries, 1956-1987
Dummies reflecting time elapsed since benchmark year, forward-looking tax burden measure, con-trolling for home taxes and home country double taxation relief
BL, FL OLS No tax effect for retained earnings, while transfers showed significant reac-tions. No clear insights about the influence of home country taxation.
Cassou (1997) Bilateral US inbound FDI from 7 countries, 1970 - 1989
Time series analysis, then extend-ing Slemrod (1990) to panel data analysis, but no control for double taxation relief system
BL OLS, home country dummies (when panel)
A significantly negative impact of the corporate tax rate on inward FDI was identified.
Tax measures: Backward looking tax measures (BL), Forward looking tax measures (FL) Source: Own Compilation.
2.2 Panel Data Analysis on Aggregate FDI
The idea of using panel data on bilateral FDI flows goes back to Devereux and Freeman
(1995) (see Table 2). Their sample included 7 OECD countries in the period from 1985 to
1989. As a further innovation, Devereux and Freeman (1995) introduced bilateral cost of
capital into their analysis reflecting the country-pair-specific effective marginal tax burden on
- 7 -
transnational investments. They found that location choice is indeed affected by taxation.
Since then, bilateral FDI flows or positions have remained a commonly used basis for subse-
quent research.6 For a long time, Buettner (2002) is the only one using bilateral cost of capital
while most of the other studies based on bilateral FDI rely on unilateral tax measures.7 Draw-
ing on a panel of intra-EU transfers of funds between 1991 and 1998, Buettner also includes
statutory corporate income tax rates into his regressions thus attempting to catch some effects
of infra-marginal, i.e. localization-relevant, considerations. From a (joint) significance of
both, the effective marginal tax burden (EMTR) and the statutory tax rate (STR), he infers an
impact of the effective average tax rate (EATR) in the determination of FDI. Another note-
worthy result is that explicitly controlling for the supply of public services does not reveal an
important impact of public expenditures on locational attractiveness.
Gorter and Parikh (2003) collect OECD panel data of bilateral FDI positions. While they do
not analyze the joint effect of distinct tax measures, they separately analyze the effects of the
unilateral host country EMTR and of a backward-looking average tax rate computed on the
basis of micro data. The tax effects are significantly negative, but provide ambiguous evi-
dence for the home country system of international double taxation relief by means of a re-
stricted SUR model. While restricted regressions that employ the micro average tax rate (Mi-
cro ATR) substantiate zero tax elasticity for credit countries (U.K.), the corresponding EMTR
regressions suggest no particular relief system-specific impact on FDI. Stöwhase (2005a)
draws on bilateral FDI from 8 EU countries to Germany, the Netherlands and the UK between
1995 and 1999. He shows that only the secondary and tertiary sectors are sensitive to taxation
while the primary sector is not. Bénassy-Quéré et al. (2005) employ various, i.e. forward-
looking, backward-looking and statutory tax measures without revealing fundamental differ-
ences in the implied tax sensitivity of FDI. The effects they calculate on the basis of their
panel of intra-OECD bilateral FDI flows between 1984 and 2000 are generally negative and
significant. Interestingly, by controlling for the amount and the composition of public expend-
itures, Bénassy-Quéré et al. (2005) find a positive composition effect, i.e. a higher share of
investment in public expenditures adds to locational attractiveness and weakens tax effects.
Bellak et al. (2007) even make an attempt to test Tiebout’s intuition. Based on a panel of bila-
teral outward FDI flows of seven home countries to eight Central and Eastern European
Countries (CEEC) for the period from 1995 to 2004, they explicitly control for a possible in- 6. Panel studies using aggregate non-bilateral data are Jun (1994), Swenson (1994), Billington (1999), Broek-
man and van Vliet (2001). 7. Devereux and Freeman (1995) provide the bilateral costs of capital employed by Buettner (2002).
- 8 -
terrelation between bilateral effective average tax rates (BEATR) and infrastructure as deter-
minants of FDI. Their results provide evidence that infrastructure – especially a good endow-
ment of communication facilities – positively influences FDI; this is directed towards the re-
spective country and may even compensate for higher taxes. Demekas et al. (2007) also in-
clude a proxy for infrastructure into their study. Their panel analysis of aggregate FDI-flows
between 16 home countries and 24 CEECs during the period 1995 – 2003, which is not meant
to finance privatization-related projects, employs the statutory corporate income tax rate as a
measure of company tax burden. Based on this data set, they conclude that infrastructure de-
termines the attractiveness of developing countries, although this is not valid for countries that
have reached a higher stage of economic development. Moreover, their reported tax effects
are mainly significant. Bénassy-Quéré et al. (2005) examine US foreign investment in 18 EU
countries from 1994 – 2003. They assume a trade-off between public spending and the tax
rate; public spending as well as a lower tax rate can increase FDI. Their findings confirm their
assumption. They emphasize that it is not desirable to participate in fierce tax competition
because of negative spillovers, such as lost funding for public goods. However, they argue
that tax competition leads to a relocation to more efficient public goods. Wijeweera et al.
(2007) employ an OECD panel data set of FDI from nine source countries investing in the
U.S. during the period from 1982 to 2000. They include various controls into their regressions
which separately resort to forward-looking marginal as well as infra-marginal tax measures as
well as to the statutory tax rate. The measured tax effects are more robust for the statutory tax
rates as for the effective tax measures. Controlling for infrastructure yields mainly insignifi-
cant coefficients, which are explained by the generally high level of economic development
and, thus, infrastructure endowment of all home countries included in the panel.
Bellak and Leibrecht (2007) concentrate on Central and Eastern Europe as the target for FDI
flows. They use panel data on bilateral FDI flows from seven home countries to eight CEECs
for the time span from 1995 to 2003. The authors integrate both, EMTR as well as the EATR,
into their analysis. Furthermore, like Bellak, Leibrecht and Damijan (2007), they belong to
the relatively few studies that do not rely on simple unilateral measures which only reflect
host country specific tax legislation explicitly referring to bilateral measures. Their results
indicate that FDI-flows to CEECs are primarily determined by bilateral effective tax rates
(BEATR). Bellak and Leibrecht (2007) argue that this is due to the predominance of new and
inframarginal investment projects in these transition economies. Hajkova et al. (2006) use
BEATR in their regressions drawing on a panel of bilateral intra-OECD FDI-stocks over the
1990s. They apply the transformed least squares approach introduced by Erkel-Rousse and
- 9 -
Mirza (2002) which allows to control for unobserved heterogeneity specific for each home
country, host country and home-host country pair. Generally, they report rather small tax elas-
ticities. The authors explain this result by dint of their very broad set of policy controls. Egger
et al. (2008) particularly focus on the interrelation of bilateral and unilateral effective tax
rates. Based on a panel of outbound positions of bilateral intra-OECD FDI between 1991 and
2002, Egger et al. (2008) show that the estimated impact of the host country unilateral tax rate
turns positive, when the bilateral tax rate is additionally considered. This reflects the ceteris-
paribus-advantage of multinationals over competing domestic firms in case of rising unilateral
host country tax rates. The coefficient of the bilateral tax rates, however, is robustly negative
as it translates the direct negative impact on after-tax returns of transnational investments.
Table 2: Empirical studies using aggregate panel data
Study Data Special Remarks Tax Measure Method Main results
Jun (1994) US aggregate in-bound FDI from OECD countries, 1980-1989
Controlling for home country system of double taxation relief
BL, FL Fixed effects : OLS, GLS, country dummies
The home country’s statutory tax rate has a significant negative effect on FDI from countries with residence systems, but not on FDI from terri-torial systems.
Swenson (1994)
Bilateral US inbound PPE investments, 1979-1991
Exploiting US tax history of the 1980s, focusing on industry variation (18 industries)
BL, FL OLS, GLS, industry dum-mies, time trend
Foreign investment in the US is tax sensitive. Average tax rates are more precise indicators for the effects than effective tax rate measures.
Devereux and Free-man (1995)
Bilateral FDI, 7 OECD coun-tries, 1985-1989
Bilateral effective tax burden measure
FL Fixed Effects: IV, OLS, time and country-pair dummies or FD estimator
Location choice is affected by taxa-tion. Granting tax credit for foreign shareholders may lead to an increase of inbound FDI from exemption countries.
Billington (1999) Aggregate in-bound FDI, G7-countries, 11 UK regions, 1986 - 1983
STR OLS The statutory corporate tax rate has a significantly negative effect on FDI inflow.
Shang-Jin (1997) Bilateral FDI, from 14 source to 46 host countries, 1990-1991
Explicitly control-ling for corruption
BL (Quasi-)Fixed Effects: OLS, IV, source country dummies
A rise of the tax rate as well as a rise in corruption leads to a reduction of inward FDI. U.S. investors are not significantly more averse to corrup-tion than are average OECD-investors.
Broekman and Vliet (2001)
Aggregate EU inbound FDI, 1989-1998
BL Fixed Effects: OLS, time dum-mies, host coun-try dummies.
There is a significant relation between effective tax rates and FDI with a tax elasticity of 0.5.
Buettner (2002) Bilateral FDI financed by transfers, EU 15, 1991-1998
Effects of statutory and effective mar-ginal tax burden analysed jointly to capture average effective tax bur-den. Controlling for public services.
BL, FL, STR
Fixed Effects: OLS, IV, home and host country dummies or FE estimator. Random Effects.Always: Time dummies.
The effective average tax burden significantly determines FDI. Public services do not seem to have an important impact on locational attrac-tiveness.
Gorter and Parikh (2003)
Bilateral FDI positions,
EU 15, 1995-1996
Extending Hines (1996)
BL, FL OLS, SUR Tax effects on FDI are significantly negative. Accounting for the home country system of international double taxation relief provides ambiguous evidence.
- 10 -
Table 2 (cont.): Empirical studies aggregate panel data
Study Data Special Remarks Tax Measure
Method Main results
Bénassy-Quéré et al. (2005)
Bilateral FDI, 11 OECD countries, 1984 - 2000
Various specifica-tions, controls and tax burden meas-ures.
BL, FL, STR
Fixed Effects: OLS, IV, home and host country dummies
Tax effects are generally significantly negative. A higher share of invest-ment in public expenditures weakens tax effects.
Razin et al. (2005) Bilateral FDI, 24 OECD countries, 1981 - 1998
2-stage investment decision analyzed with macroeconom-ic data
STR Heckman selec-tion: Probit, OLS, time and country dummies
Source-country statutory tax rate plays an important role in the selec-tion process, whereas the host country tax rate affects the scale of invest-ment.
Stöwhase (2005a) Bilateral FDI, from 8 EU coun-tries to Germany, Netherlands and UK, 1995-1999
Investments diffe-rentiated by three main economic sectors
BL, FL Fixed Effects: GLS, SUR, country-pair dummies and time dummies.
Tax sensitivity of FDI varies between economic sectors. Only the secondary and tertiary sectors are sensitive to high tax rates.
Haijkova et al. (2006) Bilateral FDI stocks, 28 OECD countries, 1991 - 2000
Fully controlling for unobserved heterogeneity
FL, STR Fixed Effects: FE estimator, Trans-formed Least Squares (TLS)
Generally rather small tax elasticities are reported, presumably due to a very broad set of policy controls.
Bellak and Leibrecht (2007)
Bilateral FDI, 7 home to 8 CEE countries, 1995 - 2003
Bilateral forward-looking tax burden measures
FL Random Effects, time dummies. Fixed Effects: FE Estimator.
FDI into CEECs is primarily deter-mined by the bilateral effective aver-age tax burden.
Bellak et al. (2007) Bilateral FDI, 7 home to 8 CEE countries, 1995 - 2004
Controlling for infrastructure, bilateral forward-looking tax burden measures
FL Random Effects, time dummies
Infrastructure - especially communi-cation facilities - positively influences FDI and may compensate for higher taxes.
Bénassy-Quéré et al. (2007)
Bilateral US outbound PPE stock in 18 EU countries, 1994-1999 or 2003
Testing if the provision of public goods changes the tax-sensitivity of FDI
FL, STR Fixed Effects: OLS, country and/or sector dummies, alter-natively FD Estimation
The provision of public goods does affect location choice. The increase of public capital by means of higher tax rate though can have negative effects on inward FDI-flows.
Wijeweera et al. (2007)
Bilateral US inbound FDI from 9 OECD countries, 1982 - 2000
Extending Slemrod (1990) to panel data analysis
FL, STR Fixed Effects: OLS, GLS, country dum-mies, time trend
Tax effects are more robust for the statutory tax rates as for the effective tax measures. Coefficients for infra-structure are mainly insignificant, presumably due to the high level of economic development of all coun-tries considered.
Demekas et al. (2007) Aggregate FDI, 16 home to 24 CEE countries, 1995 – 2003
Threshold effect estimation with respect to income levels, non-privatization FDI
STR Fixed Effects: S- GMM, time dummies
Apart from gravity factors, FDI in European transition economies is influenced by a broad set of host country policies.
Wolff (2007) Bilateral FDI, EU 27 (excl. Romania), 1994 - 2003
Extending Razin et al. (2005), distin-guishing total FDI, equity and debt transfers
STR Heckman selec-tion: Probit, OLS, time and country dummies
Results are mixed: For total and equity FDI no robustly significant tax effects are reported.
Egger et al. (2008) Bilateral FDI, 22 source to 26 host countries (OECD), 1991 - 2002
Focus on the inter-relation of uni- and bilateral forward-looking tax burden measures
FL Fixed Effects: FE estimator, GMM, time dummies
FDI reacts positively to the parent and host country tax rate and negatively to bilateral effective tax rate. Isolated unilateral tax rates can lead to biases results.
Tax measures: Backward looking tax measures (BL), Forward looking tax measures (FL), Statutory tax rates (STR) Source: Own Compilation.
- 11 -
As Devereux (2006) points out, FDI-flows (or the resulting positions) rather describe one me-
thod to finance expansion and do not correspond precisely to any specific stage of the overall
investment decision. However, Razin et al. (2005) apply the idea of two sequential decisions
– how much to invest conditional on the discrete decision to invest at all – even to bilateral
macroeconomic FDI data. They analyze bilateral panel data with 24 OECD countries over the
period 1981 to 1998 and address the underlying sample selection problem by the Heckman
selection method. Thereby, they do not focus on marginal versus infra-marginal effective tax
measures, but on home and host country STRs. Their results indicate that the source-country
statutory tax rate is an important factor in the selection process, whereas the host-country tax
rate affects the scale of investment. Wolff (2007) closely follows Razin et al. (2005) in his
empirical approach. He draws on a panel of bilateral FDI flows for the EU 27 countries (ex-
cluding Romania) for the period 1994 – 2003 and runs regressions for total FDI as well as
equity FDI, inter-company debt transactions and retained earnings. His results are mixed. For
total and equity FDI no robust significant tax effects are reported.
2.3 Studies Using Aggregate Data on Affiliates of Multinational Companies
2.3.1 Location Choice: Count Data Models
Papke (1991) applies a Poisson count model estimated via maximum likelihood techniques in
order to analyze the tax sensitivity of location choices (see Table 3). Count data models imply
the aggregation of the number of considered companies’ affiliates per host country. Papke
(1991) draws on a panel of plant birth data built from cross-sections of the US Establishment
and Enterprise Microdata File (USEEM) for the years 1976, 1978, 1980 and 1982. Thereby,
she is among the first to use a forward-looking tax burden measure. Her results show that
economic factors such as taxes play a significant role in the determination of manufacturing
locations, although responsiveness differs considerably across industries. Stöwhase (2005b)
also relies on count data, but rejects the Poisson model in favor of a negative Binomial model
due to less restrictive assumptions. His data set is based on data covering the foreign engage-
ments of German multinationals for the period 1991 to 1998. He shows that real activity is
significantly deterred by rising effective tax burdens, while management and finance as well
as research and development (R&D) activities only react negatively to statutory tax rates.
Overesch and Wamser (2008a) examine MiDi-data for the time from 1996 to 2005 and focus
on German investments in eastern European EU accession countries. The Poisson count mod-
el proves to be applicable for the analysis of location decisions. The negative Binomial model
is only a viable alternative if country fixed effects are included. Under both models the STR
- 12 -
exerts a significant impact on the expected number of German affiliates in a certain host coun-
try. Moreover, the EATR proves to be the more suitable indicator for tax effects in contrast to
the EMTR. Controlling for infrastructure yields significant coefficients as expected in transi-
tion economies. Overesch and Wamser (2008b) use Midi-data for German outbound invest-
ments in 30 European countries for the period from 1989 to 2005 and study asymmetries with
respect to tax sensitivities originating from different motivations to internationalize, different
degrees of internationalization and different business activities of firms. On the basis of a
negative Binomial count model, they show that vertical FDI in manufacturing is significantly
more responsive to local taxation than horizontal FDI. Financial service and R&D activities
are highly tax sensitive as well as affiliates of an only weakly internationalized firm.
Table 3: Empirical studies using count data Study Data Special Remarks Tax
Measure Method Main results
Papke (1991) Data on plant births in US states, 1976 + 1978 + 1980 + 1982
Distinguishing between different industries
FL ML Estimation (Poisson count model), state fixed effects
Economic factors, e.g. taxes play a significant role in manufacturing locations, but impact differs across industries.
Stöwhase (2005b) Microdata on German multinationals’ activities in 8 EU countries, 1984 - 1994
Differentiating between functional forms of engage-ments
BL, STR
GLS (NegBin count model)
Real activities are signifi-cantly deterred by rising average tax burdens, while management, finance and R&D reacts negatively only to statutory tax rates.
Overesch and Wamser (2008a)
Microdata on German multinationals’ activities in EU 8, 1996 – 2005 (MiDi)
Inter alia controlling for infrastructure
FL, STR ML Estimation (NegBin and fixed effects: Poisson model), time dummies.
The host statutory tax rate exerts a significant impact on location choice, but the EATR is most suitable to measure tax effects. Infrastructure shows significant coefficients.
Overesch and Wamser (2008b)
Microdata on German multinationals’ activities in 30 European countries, 1989 - 2005 (MiDi)
Focus on tax asymmetries: motivation, activity, degree of internationalization
FL, STR ML Estimation (NegBin count model), time dummies
Host-country taxation has a higher impact on verti-cally than on horizontally integrated FDI.
Tax measures: Backward looking tax measures (BL), Forward looking tax measures (FL), Statutory tax rates (STR) Source: Own Compilation.
2.3.2 Continuous Capital Data: Studies Using Aggregate PPE Stocks
The early study by Grubert and Mutti (1991) relies on cross-sectional data from the Bureau of
Economic Activity (BEA), US Department of Commerce (see Table 4). These data are based
on the level of the capital (PPE) stock of affiliates of US multinational companies in specific
foreign countries. Grubert and Mutti (1991) report an important effect of taxes as well as ta-
riffs on the international capital allocation. Based on similar data with its focus on tax haven
countries, Hines and Rice (1994) find a very high negative influence of taxes on PPE posi-
tions which are higher than the estimates by Grubert and Mutti (1991). Devereux and Lock-
wood (2006) draw on the same data source as Grubert and Mutti (1991) and Hines and Rice
- 13 -
(1994) using panel data, however. Devereux and Lockwood (2006) develop a two-stage mod-
el, which is estimated in a probabilistic choice framework, in order to be able to better analyze
the two main elements of the decision-making process even on the basis of aggregate data.
The results indicate a large and significant impact of EATR, but no impact of EMTR.
Table 4: Empirical studies using aggregate PPE data on affiliates Study Data Special Remarks Tax
Measure Method Main results
Grubert and Mutti (1991)
PPE stock of US affili-ates in 33 foreign coun-tries, 1982
Considering host country taxes as well as tariffs
BL OLS Taxes and tariffs signifi-cantly determine interna-tional capital allocation.
Hines and Rice (1994)
PPE stock of US affiliates in 73 foreign locations, 1982
Special focus on tax havens BL OLS, IV A high negative influence of taxes on PPE positions is reported, exceeding the estimates of Grubert and Mutti (1991).
Hines (1996) US state-specific PPE stock of affiliates from 7 foreign countries, 1987
Central Idea: Non-uniform distribution of activities across states from credit countries captures all non-tax state-specific factors
STR Tobit There are large negative tax effects in the determi-nation of foreign direct investment.
Grubert and Mutti (2000)
PPE stock of US affiliates in 60 foreign locations, 1992
BL OLS Significantly negative tax elasticities are reported.
Altshuler et al. (2001)
PPE stock of US affiliates in 58 foreign locations, 1984+1992
Development of tax sensi-tivity from 1980s -1990s
BL, STR OLS, IV Capital has become more responsive to taxes during the 1980s.
Devereux and Lock-wood (2006)
PPE stock of US affili-ates in 19 OECD coun-tries, 1983 - 1998
2-stage model integrating probabilistic choice frame-work
FL, STR Fixed Effects : GMM, time and country dummies
Results indicate a large and significant impact of the EATR, but no impact of the EMTR.
Bobonis and Shatz (2007)
US state-specific PPE stock of affiliates from 7 foreign countries, 1977-1996
Focus on agglomeration effects
STR Fixed Effects: OLS, state-home dummies, GMM
The insights as to the influence of corporate taxes remain only vague. In contrast, agglomeration effects are confirmed.
Tax measures: Backward looking tax measures (BL), Forward looking tax measures (FL), Statutory tax rates (STR) Source: Own Compilation.
Basically, Grubert and Mutti (2000) as well as Altshuler et al. (2001) have access to individu-
al company tax-returns data. However, they aggregate this information on US outbound activ-
ities within each host country. Grubert and Mutti (2000) regress the amount of invested real
capital by the foreign affiliates of US multinationals on ATR in about 60 host countries. They
report significantly negative tax elasticities. Altshuler et al. (2001) use similar data and find
out that capital has become more responsive to taxes during the 1980s. Hines (1996) uses
BEA data as well, but on the US activities of foreign multinationals. He analyzes variations in
the distribution of PPE between US states and identifies non-tax factors by looking at the dis-
tribution of investments from tax credit countries, restricting their tax sensitivity to zero. For
exemption countries, Hines (1996) reports large and negative tax effects in the determination
of FDI. Hines’ approach has been adopted by following authors such as Gorter and Parikh
(2003). Bobonis and Shatz (2007) use the same BEA data source as Hines (1996) on the US
state-level foreign-invested capital stock. But similar to Devereux and Lockwood (2006), they
- 14 -
rely on a panel covering the years from 1977 to 1996. Bobonis and Shatz (2007) resort to a
dynamic specification and especially focus on agglomeration effects. The existence of such
effects on PPE positions is confirmed. However, the insights about tax effects remain vague.
2.4 Studies Using Individual Company Data
While economists from the US have access to individual data since several years, comparable
data for Europe has become available only recently. The improved availability of micro data
has led to a rising number of studies, which focus on the distinct levels of the overall foreign
investment decision.
2.4.1 Location Choice: Discrete Choice Models
Bartik (1985) was the first to employ the (nested) conditional Logit model developed by
McFadden (1974, 1978) on microeconomic company data (see Table 5). His aim was to dis-
cern the influencing factors of discrete location choices between US states. The information
on plant locations stems from a data base developed by Schmenner (1982) for the years 1972
and 1978. He reports a significant negative effect of state taxes on business location patterns.
Devereux and Griffith (1998) also use a nested conditional Logit model to analyze compa-
nies’ discrete decision where to produce. They find evidence that the EATR – as opposed to
the EMTR – plays a significant deterring role in the location decision. This result was found
on the basis of a Compustat panel for the period 1980 to 1994 covering US firms that are lo-
cated in the European market. Swenson (2001) estimates a conditional Logit model as well,
using panel data on discrete investment choices that were made by multinational companies
from seven countries in different states of the US between 1984 and 1994. Her data set allows
for distinguishing between alternative transaction types and the source country system of
double taxation relief. With mergers and acquisitions (M&A) being positively affected by
rising host country taxes, some of her results, however, turn out to be rather surprising.
Mutti and Grubert (2004) employ a random effects Probit analysis based on cross-sectional
individual tax returns of US multinationals investing in 60 foreign locations for the year 1996.
Controlling for parent characteristics they find that especially labor intensity positively affects
international mobility. At the policy level taxation is an important factor. However, its ad-
verse effect on location decisions is weakened in high-income host countries, probably due to
better infrastructure or agglomeration effects there. Buettner and Ruf (2007) opt for a fixed
effects Logit model, which is based on more realistic assumptions as compared to previous
studies. For example, it allows parent companies to hold more than one subsidiary abroad.
- 15 -
Another virtue of this approach is that (time-invariant) heterogeneity in the firms’ assessment
of locations and heterogeneity of locations itself can be addressed simultaneously by means of
fixed firm-specific location effects. They base their data-set on a panel of German multina-
tionals that invest in 18 OECD countries; the panel is provided by the German Bundesbank’s
MiDi database and covers the time span from 1996 to 2003. Their results confirm the results
of Devereux and Griffith (1998) regarding the irrelevance of the EMTR in discrete location
decisions. However, in contrast to Devereux and Griffith, STR shows a stronger adverse ef-
fect on location choice as compared to the EATR, which is only weakly significant. Buettner
and Ruf (2007) hypothesize that the latter result might reflect measurement problems, i.e. an
imperfect fit of the employed effective tax rates to individual firms and investments. Buettner
and Wamser (2008) also examine the discrete location decision by means of a fixed effects
Logit model drawing on MiDi-data for 22 OECD host countries from 1996 to 2004. They
report robust negative effects on the parent-specific probability to invest for the statutory cor-
porate tax rate in accordance with Buettner and Ruf (2007). However, Buettner and Wamser
(2008) consider several non-profit taxes, which mostly do not show significant effects.
Table 5: Empirical studies applying discrete choice models on micro data Study Data Notes Tax Method Main results
Bartik (1985) Data on plant births in US states, 1972 + 1978
Numerous controls STR Cond. Logit There is a significant negative effect of state taxes on business location.
Devereux and Grif-fith (1998)
Microdata on US multinationals’ activi-ties in F, GER, UK, 1980 - 1994
Focus on relevance of inframarginal vs. marginal effective tax burden
FL, BL Cond. Logit, country and time fixed effects
Results show that the EATR as op-posed to the EMTR plays a significant deterring role in location decisions.
Swenson (2001) Microdata on foreign multinationals’ activi-ties in US states, 1984 -1994
Different transac-tion types, credit vs. exemption countries
STR Cond. Logit, state fixed effects
The tax-sensitivity of FDI differs according to transaction type. M&A are positively correlated with host country taxes.
Mutti and Grubert (2004)
Microdata on US multinationals’ activi-ties in 60 locations, 1996
Parent characteris-tics, different indus-tries
BL Random Effects Probit
Industries with above-average export-orientation are more sensitive to host country taxation.
Buettner and Ruf (2007)
Microdata on German multinationals’ activi-ties in 18 OECD coun-tries, 1996 - 2003 (MiDi)
Unobserved firm-specific location effects
FL, STR Fixed Effects Logit, fixed firm-country and time effects
Results of Devereux and Griffith (1998) are generally confirmed. How-ever, the statutory tax rate shows stronger adverse effects as the only weakly significant EATR.
Buettner and Wamser (2008)
Microdata on German multinationals’ activi-ties in 22 OECD coun-tries, 1996 - 2004 (MiDi)
Following Buettner and Ruf (2007), focus on non-profit taxes
BL, FL, STR
Fixed Effects Logit, fixed firm-country and time effects
There are robust negative effects on the parent-specific probability to invest for the statutory corporate tax rate.
Tax measures: Backward looking tax measures (BL), Forward looking tax measures (FL), Statutory tax rates (STR) Source: Own Compilation.
2.4.2 Continuous Capital Data: Scale of Investment Conditional on Location Choice
Desai et al. (2004) study the continuous investment decision conditional on locating in a cer-
tain country by means of firm-level data (see Table 6). They do not only consider corporate
- 16 -
income taxes but additionally include local indirect taxes into their analysis. Based on BEA
micro-data on US outbound activities in 1982, 1989 and 1994 they find a tax elasticity of for-
eign affiliate assets of about 7 for both indirect as well as income taxes. Overesch and Wams-
er (2008a) examine whether tax effects influence the decision on the scale of investment, i.e.
they do not confine their analysis to discrete location choices as already described above.
They show that tax effects are robust and highly significant based on their set of firm-level
data (MiDi) that covers German investments in Eastern Europe and relies on a dynamic
framework with differenced GMM estimation. This result holds independent from the use of
statutory taxes or forward-looking tax measures (EATR and EMTR). Buettner and Wamser
(2008) analyze the subsequent stages of the transnational investment decision separately. As
regards the continuous investment decision – based on a dynamic specification – they report
significant effects of their forward-looking tax burden measure which represents the tax
wedge that was introduced by corporate taxation as part of the user cost of capital. Indirect
taxes, which are additionally employed as regressors, do not reveal significant effects when
the study controls for country fixed effects.
Table 6: Empirical micro data studies analyzing continuous investment decisions Study Data Notes Tax Method Main results
Desai et al. (2004) Micro-data on US multi-nationals’ outbound activities, 1982+1989+1994 (BEA)
Additional focus on indirect taxes
BL Fixed effects: OLS, parent and industry fixed effects
Tax elasticity ranges around 7 for corporate income as well as indirect taxes.
Overesch and Wams-er (2008a)
Micro-data on German multinationals’ activities in EU 8, 1996 – 2005 (MiDi)
Inter alia controlling for infrastructure
FL, STR Fixed Effects: GMM, time dum-mies
Tax effects are robust and highly significant.
Buettner and Wamser (2008)
Micro-data on German multinationals’ activities in 22 OECD countries, 1996 – 2004 (MiDi)
Following Buettner and Ruf (2007), focus on non-profit taxes
BL, FL, STR
Fixed Effects: OLS, IV, with parent, industry and country dummies, GMM, always time dum-mies
Continuous investment deci-sions are influenced by the tax wedge. Indirect taxes do not matter as soon as country fixed effects are controlled for.
Tax measures: Backward looking tax measures (BL), Forward looking tax measures (FL), Statutory tax rates (STR) Source: Own Compilation.
3. The Meta-Dataset
3.1 Studies Covered
Our data set results from the combination of the 25 studies that are provided by De Mooij and
Ederveen (2003) and 21 additional studies.8 Six of these 21 studies, which are new to our data
set as compared to De Mooij and Ederveen (2003), have already been included in De Mooij
and Ederveen (2005, 2006), but we collected the corresponding information ourselves. Since
we exclude an earlier version of Bénassy-Quéré et al. (2005) from the old data set taken from
8. Many thanks again to Ruud de Mooij for providing his data to us.
- 17 -
De Mooij and Ederveen (2003), we finally obtain a sample of 45 studies of which 15 (33 %)
have never been meta-analyzed before.9 Thus, all studies, which appeared in our review of the
empirical evidence in Section 2, are covered by our data sample.
3.2 New Variables
In addition to the inclusion of new studies we added some new study or model characteristics
to the original data set assembled by De Mooij and Ederveen (2003) which have not been
coded before. In particular, we tried to capture model specifications in more detail. We there-
fore coded two dummies which respectively take a value of one if primary studies controlled
for country or time fixed effects and zero otherwise. As regards tax measures employed in
primary work, we now distinguish whether effective average or marginal tax rates reflect only
host country provisions or also enclose bilateral tax agreements. For the latter case we code
two new variables named bilateral effective average tax rate (BEATR) and bilateral effective
marginal tax rate (BEMTR). In addition, we also introduce dummy variables adopting a value
of one if the primary model controlled for some measure of public spending relevant from an
investors’ point of view (such as public inputs, public investment, infrastructure or educa-
tion). Likewise, agglomeration effects are captured. Finally, we more closely describe the
properties of the data by introducing dummies for studies that are based on aggregate data,
aggregate data on affiliates or firm-level data, respectively. Another modification as compared
to the old dataset concerns the classification of studies according to the properties of the data.
We classify each study as either being based on time series, cross-section or panel data and
consider the question whether discrete location choice has been examined as constituting a
separate dimension of the literature. In particular this implies a deviation from the methodolo-
gy chosen by De Mooij and Ederveen (2005, 2006) who treat studies that analyze discrete
location choice as an alternative to time series, cross-section or panel data studies. As a result,
we are able to precisely capture the different dimension of the primary literature as outlined in
Section 2 within the scope of our meta-analysis.
3.3 Meta-Samples
We employ two different measures of primary effect size estimates. On the one hand, we rely
on semi-elasticities, which are computed on the basis of the original primary estimates. The
transformation of original estimates into comparable semi-elasticities (evaluated at the sample
means) is already done by De Mooij and Ederveen (2003) for the observations included in
9. The excluded study is Bénassy-Quéré et al. (2001).
- 18 -
their meta-study. Whenever necessary, we apply this procedure to those studies which have
been newly adopted for our meta-analysis.10 For some studies or models, however, no semi-
elasticities could be calculated. In particular, this holds for Buettner and Ruf (2007) as well as
some models in Buettner and Wamser (2008). In these cases fixed effects logit estimation is
applied which does not allow for the computation of the marginal effects needed to derive the
tax semi-elasticities of foreign investment decisions. In addition to semi-elasticities, we also
will employ primary t- and z-values as dependent variable in the meta-analysis. As is com-
monly known, t- and z-values result from the scaling of primary coefficients by their respec-
tive standard errors.11 They thus are by definition standardized and defined on a dimension-
less scale. Consequently, they are comparable across different studies and models. However,
some primary models include non-linearities with respect to the relationship between FDI and
taxes. In these cases no exact standard errors of the total marginal effect can be calculated
without information on the corresponding covariances. Following De Mooij and Ederveen
(2003) we therefore resort to the assumption of zero covariances.12 As we also include z-va-
lues obtained from fixed effects Logit estimates in Buettner and Ruf (2007) and Buettner and
Wamser (2008), the t/z-value meta-sample will include slightly more observations as com-
pared to the sample for the analysis based on semi-elasticities. Furthermore, note that we ex-
clude observations from both samples if the value of the semi-elasticity does not lie within a
range of twice the standard deviation from the mean semi-elasticity of the study. Also note
that in analogy to De Mooij and Ederveen (2003, 2005, 2006) we premultiply each semi-
elasticity and each t-value by (-1). Thus, the higher one of these effect size measures, the
higher is the implied adverse tax effect.
4. Some Descriptive Statistics
As Table 7 shows, the meta-dataset contains 730 observations in total. 35.1% of these obser-
vations are obtained from panel analyses on aggregate continuous capital data. Around 24%
of the observations are taken from micro data studies (study classes 1, 2 and 3). Studies using
aggregate data on affiliates (incl. count data) make up around 18% of the observations used in
the meta-analysis. Time series studies mainly published in the 1980s still make up 13% of the
meta-dataset. The ratio of published work is generally high within the different study catego- 10. Many primary models, i.e. linear models with a log-level specification, directly produce point estimates of
semi-elasticities. Hence no transformation is necessary. 11. For convenience we will only refer to “t-values” instead of “t- and z-values” subsequently. 12. An alternative procedure would be to exclude the respective observations from the data sample underlying
our t- and z-value based analysis. However, this would imply that the analysis based on semi-elasticities and the analysis based on the t-and z-values would rely on different data samples. This again would significantly reduce their comparability.
- 19 -
ries. However, especially count data models exhibit until today a rather low publication ratio.
This is explained by the relatively recent accomplishment of some of this research.
Table 7: Descriptive statistics Study
Class
Observations (abs./ %)
Semi-
elasticities
T-Values Published (abs./ %)
Tax type (abs./ %)
mean sd. mean sd. Backward Forward Statutory
1 12 1.6% 0.2 0.8 5.0 3.9 12 100.0% 12 100.0% 0 0.0% 0 0.0%
2 40 5.5% 1.0 0.5 3.7 2.2 31 77.5% 3 7.5% 33 82.5% 4 10.0%
3 122 16.7% 2.0* 5.0* 1.1 2.1 122 100.0% 2 1.6% 19 15.6% 101 82.8%
4 73 10.0% 6.0 4.8 2.7 1.0 73 100.0% 39 53.4% 0 0.0% 34 46.6%
5 12 1.6% 4.2 3.5 1.7 1.4 12 100.0% 1 8.3% 0 0.0% 11 91.7%
6 21 2.9% 1.2 4.6 0.9 1.4 14 66.7% 0 0.0% 6 28.6% 15 71.4%
7 95 13.0% 2.2 2.6 1.8 1.4 15 15.8% 11 11.6% 50 52.6% 34 35.8%
8 95 13.0% 2.5 4.4 1.5 2.2 95 100.0% 48 50.5% 47 49.5% 0 0.0%
9 4 0.5% 5.7 6.2 1.4 1.5 4 100.0% 0 0.0% 0 0.0% 4 100.0%
10 256 35.1% 2.9 3.2 3.3 3.9 131 51.2% 54 21.1% 139 54.3% 63 24.6%
All 730 100.0% 2.8 3.9 2.4 3.0 498 68.2% 170 23.3% 294 40.3% 266 36.4% Study Class: 1 – cross-section/micro/location choice; 2 – panel/micro/cont. capital; 3 – panel/micro/location choice; 4 – cross-section/aggregate affiliate/cont. capital; 5 – cross-section/aggregate affiliate/location choice; 6 – panel/aggregate affiliate/cont. capital; 7 - panel/aggregate affiliate/location choice; 8 – time series/aggregate/cont. capital; 9 – cross-section/aggregate/cont. capital; 10 – pan-el/aggregate/cont. capital. Abbreviations: sd.: Standard deviation. abs. : absolute numbers. Backward: Backward-looking tax rates. Forward: Forward-looking tax rates. Statutory: Statutory Tax Rates.
Strikingly, micro data studies show on average relatively low semi-elasticities while the mean
t-values are high as compared to studies based on aggregate data. Interestingly, each type of
tax burden measure – backward-looking, forward-looking or statutory tax rates – is applied in
nearly all study classes. There are, however, some exceptions. For example, forward-looking
tax rates are not at all applied in cross-sectional studies based on aggregate data on affiliates
which are actually dominated by US literature. In return, forward-looking tax rates make up
the highest share of models in panel studies based on firm-level continuous capital data.
5. A Full-Scale Meta-Study
5.1 The Methodological Framework of Meta-Analytic Estimators
Sophisticated meta-studies may resort to a broad range of meta-analytical techniques to assess
heterogeneous empirical evidence on a certain question of interest. These techniques, howev-
er, should not be considered in isolation from each other. Quite the contrary, if thoroughly
applied all meta-analytical tools combine to form a consistent methodological framework
which systematically exploits the heterogeneity of primary evidence. Meta-analytical eco-
- 20 -
nomic research explicitly focusing on conceptual issues and therefore choosing a rather tech-
nical approach to the methodology has multiplied during recent years. Instructive examples
are e.g. Stanley (2005, 2008), Brons, Nijkamp, Pels and Rietveld (2008) and in particular
Brons (2006) and Bom and Ligthart (2008).13 We follow these convincing approaches. More-
over, we present the full set of meta-analytical estimators in a formally integrated framework.
This framework then forms the basis of our meta-analytical estimation strategy.
5.1.1 Classical Meta-Analysis
In an initial step, meta-analytical techniques are used to combine single study results to an
overall estimate which subsequently can be tested for significance. One possible way to di-
rectly uncover the overall statistical significance of the relationship between host country cor-
porate taxation and foreign direct investment would be to perform Fisher’s combined test on
p-values of the individual estimates. However, this test suffers from several shortcomings. In
particular it does not indicate the sign of the effect that was tested for significance (Jarrell and
Stanley 2004). Therefore we concentrate on the computation of pooled effect estimates via the
use of fixed or random effects meta-analysis.
Conceptually, both fixed and random effects meta-analysis estimates result from weighted
least squares (WLS) regression of individual coefficient sizes on an intercept only, where the
inverse of the effect size variance is used as analytical weight.14 The fixed effects meta-
analysis builds on the rather strong assumption that all considered studies generally estimate
the same true underlying effect. Any deviation of the single effect estimate from this true ef-
fect is thus explained by pure sampling error.
Let iγ be the study-specific estimate of the true effect iγ underlying the primary studies
1,...,=i N . If the true effect is supposed to be the same for all primary studies 1,...,=i N such
that 0iγ γ= , we can write the model underlying FE meta-analysis as
0i iγ γ ε= + (1)
iε represents the error term satisfying ( )( ) 0= =i i iE Eε ε γ , and ( ) ( )2 ˆ= =i i i iV Vεσ ε γ γ . The
fixed effects pooled estimator, i.e. the WLS estimator γ of 0γ , is given by
13. The very foundations of meta-analysis, also in a technical sense, have meanwhile been laid by inter alia
Hedges and Olkin (1985), Hunter and Schmidt (1990), Stanley and Jarrell (1989, 1998), Jarrell and Stanley (1990) and Stanley (1998, 2000a, b).
14. Therefore, they should not be confused with the homonymous panel econometric approaches.
- 21 -
1
1
ˆN
i ii
N
ii
w
w
γγ =
=
= ∑
∑ (2)
As iγ is regressed on a constant only, the (efficient) WLS analytic weights are just the reci-
procal unconditional variances of iγ . Since 0iγ γ= is assumed, the unconditional observation
variance equals the variance conditional on iγ , i.e. ( ) 2ˆ ˆ( ) = =ii i iV V εγ γ γ σ . The WLS analyti-
cal weights thus simplify to 2ˆ1/ ( ) 1= =ii iw V εγ σ . By definition ( )i iV γ γ – or better: its con-
sistent estimate – is generally derived from the estimated variances matrices of the primary
regressions. Thus, the variance of the idiosyncratic error 2iεσ is assumed as known.
In contrast to fixed effects meta-analysis, random effects meta-analysis explicitly allows for a
distribution of primary effect size estimates beyond pure sampling error, assuming that the
underlying true effect sizes randomly differ from study to study.
Let iγ again be the estimate of the true effect iγ for studies 1,...,=i N and let iε again be the
error with the same (conditional) moments as defined above. By assumption, iγ is drawn
from a random distribution such that 0i iγ γ μ= + with 2(0, )i iid μμ σ∼ . Hence, we can write
0ˆ = + +i i iγ γ μ ε (3)
Again the pooled estimator is described by formula (2). However, the (efficient) weights
ˆ1/ ( )=i iw V γ applied to calculate γ differ in that they now take account of an additive va-
riance component 2μσ . While 2
μσ represents the between-study variance, 2iεσ is consistently
referred to as the within-variance. The resulting precision weights in the case of random ef-
fects meta-analysis are then given by ( )2 2ˆ1/ ( ) 1= = +ii iw V μ εγ σ σ . 2
iεσ is again derived from
the primary estimates. Instead, 2μσ has to be estimated in a first step.15
Since the underlying assumptions of fixed and random effects meta-analysis are mutually
exclusive, there is a need to test for their respective validity in order to finally choose the ade-
quate estimator for the data at hand. A common test for the null hypothesis of between-study
15. As regards estimation of 2
μσ , Thompson and Sharp (1999) describe various estimators which are applicable.
- 22 -
homogeneity, i.e. a between-study variance 2μσ of zero, is the Q-test. It is based on the 2χ -
distributed Q-statistic given by
L
1
i=1
1
2
2 2ˆ
ˆQ (L 1),=
=
⎛ ⎞⎜ ⎟⎝ ⎠= − ∼ −∑
∑∑
L
iL
i
i i
i i
i
ww
w
γγ χ (4)
with all variables as defined above and iw as the fixed effects weights.
As soon as the computed Q-statistic forces to reject the null of between-study homogeneity,
the simple fixed effects meta-analysis is no longer applicable. However, even classical ran-
dom effects meta-analysis remains limited in scope. It in fact takes account of the existing
heterogeneity beyond sampling error, but it does not attempt to systematically explain this
excess variability of primary effect size estimates.
5.1.2 Meta-Regression
Consequently, most modern meta-studies turn to meta-regression and thus go beyond classic-
al meta-analysis based on “empty models” by introducing relevant explanatory variables.
Since these meta-regression models imply a conceptual extension of classical random effects
meta-analyses, the underlying econometric models remain relatively straightforward. Howev-
er, the notation chosen hereafter will take account of the fact that many meta-studies follow
the suggestions by Bijmolt and Pieters (2001) to use the information from all estimated mod-
els in primary studies for meta-regressions (so-called multiple sampling).16
Depending on whether all the heterogeneity beyond sampling error can be systematically cap-
tured by the moderator variables, meta-regression analysis can be classified into fixed effects
versus mixed (also called random) effects meta-regression.
The fixed effects meta-regression model is given by
0
ˆ
= += + +
is is is
is
γ γ εγ γ i isx β z δ ,
(5)
where isγ represents the sth effect size estimate sampled from study i. 0γ is the intercept and
ix and isz are vectors with study-specific and model-specific variables respectively. isε is the 16. Multiple sampling leads to more powerful tests and more accurate estimates due to the larger underlying
sample as compared to single estimate sampling. Multiple sampling is applied inter alia in Feld, Baskaran and Schnellenbach (2007) or De Mooij and Ederveen (2003, 2005, 2006). The question whether single esti-mate or multiple sampling should be applied, however, is not beyond controversy (see e.g. Stanley 2001).
- 23 -
error term satisfying ( )( ) 0= =is is isE Eε ε γ . We further define
( ) ( )2 ˆ,= =is is is isV Vεσ ε γ γi isx z .
Correspondingly, the mixed effects meta-regression can be written as
0
ˆ
= += + + +
is is is
is is
γ γ εγ γ μi isx β z δ
(6)
Obviously some systematically unexplained heterogeneity 2(0, )∼is iid μμ σ remains. All other
variables are as defined in (5).
In traditional meta-studies both fixed effects and mixed effects meta-regression models are
estimated by WLS. Generally, the employed analytical weights isw correspond to the reci-
procal conditional effect size variance implying ( )ˆ1/ ,=is isw V γ i isx z . In the case of WLS
within a fixed effects meta-analytic framework it holds that ( ) ( )2 ˆ ˆ,= =is is is isV Vεσ γ γ γi isx z .
Remember that we can derive ( )is isV γ γ – or more precisely: the consistent estimate
( )ˆis isV γ γ – from the primary models. Furthermore, if WLS is applied to estimate a mixed
effects meta-regression model, ( )ˆ ,isV γ i isx z again contains an additive component
representing the between-variance. It holds that ( ) 2 2ˆ , = +isisV μ εγ σ σi isx z .17 As in classical
random effects meta-analysis, 2μσ must again be estimated. Instead, 2
isεσ still is considered as
known.
Particularly for mixed effects meta-regression models, however, estimators other than WLS
might be considered. Precisely, pooled OLS (POLS) as well as feasible GLS (FGLS) estima-
tion techniques are applicable. While POLS treats each observation equally without applying
any weighting scheme, FGLS again implies a weighting of observations. However, in contrast
to the WLS estimation the weighting matrix is not supposed to be (partly) known. It is rather
fully estimated on the basis of POLS residuals. POLS or FGLS might be favored by meta-
analysts if the unobserved heterogeneity beyond sampling error is perceived as a dominant
driver of the meta-regression disturbance. The disturbance in this case is simply = +is is isv μ ε ,
17. We have ( ) ( ) 2ˆ , ,= +
isis isV V εγ μ σi is i isx z x z .The between-variance by definition is constant across obser-
vations so that ( ) 2, =isVμ
μ σi isx z .
- 24 -
but without explicit decomposition into sampling error isε and heterogeneity beyond sampling
error isμ . isv must be assumed as iid-distributed with all classical moment conditions for li-
near regression models. As regards the fixed effects meta-regression model, it by definition
should be estimated via WLS since 2isεσ can be derived from the primary literature, i.e. the
applicable weights are very reliably known.
Apparently, all the different estimators explained so far basically differ in their applied
weighting schemes. The best choice among them should thus primarily be guided by efficien-
cy concerns. None of them, however, explicitly takes account of the cluster-sample characte-
ristics of the meta-dataset and the implied possibility of dependence of observations. Taking
these issues into account, not only efficiency is at stake, but – under certain conditions – even
consistency. However, if model (6) is transformed such that it shows a nested error structure,
it can be estimated with cluster-econometric estimation techniques that control for the unob-
served study-level heterogeneity. The meta-analytical literature in this case often refers to hie-
rarchical or multilevel models. As opposed to the fixed and mixed effects meta-regressions,
these multilevel models account for non-zero off-diagonal elements of the variance-
covariance-matrix, e.g. for within-study correlation resulting from the unobserved study-
specific effects. Basically, these multilevel models correspond to random effects models in
cluster-econometric terms18 and as such can be written as
0
ˆ
= += + + +
is is is
is i
γ γ εγ γ μi isx β z δ
(7)
with 2(0, )∼i iid μμ σ being the unobserved heterogeneity at the study-level.
However, random effects models (in cluster-econometric terms) should be considered with
caution. Not only should one test for the mere existence of an unobserved study-level effect,
but equally important is a test whether the nested error and the meta-regressors are uncorre-
lated. Surprisingly, neither test is routinely performed in multilevel meta-analyses. Instead,
the absence of an endogeneity problem is simply imposed by assumption. If, however, unob-
served study-level heterogeneity exists and it is correlated with the meta-regressors, multile-
vel meta-analysis (as well as pooled estimators) will not be consistent. In this case only a
fixed effects cluster-econometric approach is viable.
18. Rather basic multilevel models are also referred to as random intercept models. In principle, they might be
extended to more complex forms known as random coefficient models (see e.g. Goldstein 1995). Also more than two levels might be considered. Econometrically this affects the definition of the nested error terms.
- 25 -
5.2 Publication Bias
So far we focused on the methodological framework integrating all meta-analytical estima-
tors. However, another important issue in the meta-analytical literature is the identification
and remedy of publication bias.19 It is presumed that authors or referees of primary literature
might – even unconsciously – prefer statistically significant results over the reporting of in-
significant effects in published work. Authors might adjust their model specifications, em-
ployed estimators or data samples until significance is eventually found. If this were indeed
the case, particularly authors working on small-sample studies and therefore struggling with
imprecise estimates would have to seek to produce large effect size estimates. In this line of
reasoning, nearly all meta-analytical techniques designed to identify publication bias draw on
the correlation of primary effect size estimates and their respective standard errors.
Consequently, most statistical instruments for the detection of publication bias imply the in-
troduction of the primary effect measures’ standard error as a moderator variable into the me-
ta-analytical regression equations. A significant partial effect of the primary standard error on
effect size then provides evidence for the existence of publication bias. From a methodologi-
cal perspective, these tests for publication bias can extend classical meta-analysis as well as
sophisticated meta-regressions which already control for the influence of other relevant study
and model characteristics. Many of the statistical models coping with publication bias as e.g.
presented in Stanley (2005) are, however, somewhat restricted. Precisely, they assume that the
potential publication bias is only one-directional. Primary work then would not just seek sig-
nificance, but the effect searched for is additionally supposed to exhibit a predetermined sign.
More general models which instead allow for bi-directionality of publication bias mainly rely
on the absolute primary effect size estimates as dependent variable. They therefore might in-
deed verify the significance of even bi-directional publication bias. They are, however, not
able to simultaneously compute meaningful partial effects of model and study characteristics
on the primary effect size estimates or – with respect to classical meta-analysis – a meaning-
ful pooled effect size measure.
Only recently, Bom and Ligthart (2008) have developed a framework allowing for direction-
ally unrestricted publication bias which does not rely on the absolute effect size estimates as
dependent variable. Their suggested econometric solution works in a random as well as in a
19. For a concise methodological survey see Stanley (2005). Stanley has been pioneering meta-analysis in eco-
nomics with a focus on publication bias (see also e.g. Stanley 2001, 2008).
- 26 -
fixed effects meta-analytical context. It can by construction be applied to both classical meta-
analysis as well as meta-regressions. We will, however, only focus on the meta-regression
application of the procedure developed by Bom and Ligthart.20 This yields the following gen-
eral meta-regression formula:
( ) ( )0 1 2ˆ ˆ ˆ= + + + + +is is is pis is is nis isse D se D vγ γ ϕ γ γ ϕ γ γi isx β z δ
All parameters and variables are defined as above. The scalars 1ϕ and 2ϕ measure the (possi-
bly asymmetric) effect on isγ exerted by primary model precision as measured by the primary
estimate’s standard error ( )is isse γ γ . pisD ( )nisD are dummies equaling one if ˆ 0>isγ
( )ˆ 0<isγ . These dummies actually allow for the bi-directionality of the bias. The disturbance
term isv contains the error components and thus differs according to whether meta-regression
model (5), (6) or (7) is assumed.
Following the idea of publication bias, meta-regression models using primary t-values as de-
pendent variable should not include primary standard errors as a regressor. In the logic of
publication bias there is correlation between estimated effect size magnitude and precision
because authors with the aim to publish are persistently seeking high t-values. Consequently,
there should be no correlation between these t-values and the standard errors if publication
bias was present. What is more, such a specification would imply a severe simultaneity prob-
lem. We will, however, control for the existence of publication bias via the inclusion of a
dummy variable reflecting the publication status of primary work.
6. Estimation Strategy
6.1 Classical Meta-Analysis and First-Stage Meta-Regressions
Our estimation strategy will be based on a step-by-step procedure. First, a classical meta-
analysis will be performed. It will produce information about the presumed existence of some
heterogeneity beyond sampling error and the adequate pooled effect estimate. Subsequently,
the focus will be on the full set of meta-regression estimators outlined above. These first-
stage meta-regressions will thereby include a very basic set of meta-regressors which accord-
20. Extending classical meta-analysis with respect to publication bias actually boils down to a regression on one
moderator variable. The results of simple linear bivariate regressions, however, might be somewhat mislead-ing as they do not necessarily correspond to multiple regressions on the full set of relevant variables.
- 27 -
ing to the literature review in Section 2 are assumed to be most fundamental. Eventually, after
identifying the underlying data characteristics by means of suitable tests we will be able to
choose our preferred model. On the basis of this model which best fits the meta-data at hand
we will estimate several extensions of the basic model focusing on different research ques-
tions. Subsequently, some further robustness checks will be performed.
As to the sequence of meta-regression estimators, the simplest but most fundamental first-
stage meta-regression will be the POLS estimate of model (6). Following Wooldridge (2002,
2003), standard errors robust to arbitrary heteroscedasticity and correlation within clusters
will be calculated. The latter seems highly advisable in the case of POLS since it neglects the
possible existence of study-specific unobserved effects as well as possible correlation in the
idiosyncratic errors. In other words, we have to make POLS standard errors immune to the
possibility that model (7) rather than (6) is the true model. Leaving existing unobserved ef-
fects not modeled would produce intra-cluster correlation and imply underestimated standard
errors, thus leading to incorrect, i.e. anti-conservative, inference. Second, as regards correla-
tion in the idiosyncratic errors, Wooldridge (2003) hints at possible intra-cluster correlation if
common slopes across clusters (here: studies) are assumed.
According to the methodological system of meta-analytical estimators outlined above we will
estimate model (5) with “traditional” WLS. Besides simple POLS, model (6) will also be es-
timated by WLS as it is regularly done in traditional meta-analysis. Moreover, model (6) will
be estimated by FGLS, as well. By weighting with the inverse of the conditional variance ma-
trix - either estimated from POLS residuals (FGLS) or taken from the primary studies (WLS)
- these estimators are able to explicitly correct for the supposedly high degree of heterosce-
dasticity in the meta-regression disturbances. However, for the same reasons as in the case of
POLS we will again compute standard errors clustered at the study-level for all estimators.
Finally, the full set of meta-regression estimators will be completed with model (7) being es-
timated by random effects and the more robust fixed effects cluster-econometric techniques in
order to take account of possible study-specific unobserved heterogeneity and the implied
potential dependence problem. As regards random effects cluster-econometrics, the most
popular estimator in the meta-analytical context is the restricted maximum likelihood estima-
tor (REML). Besides REML we will also employ the standard GLS random effects estima-
tor.21
21. Both estimators are closely related as random GLS estimates can be obtained from ML estimation by re-
stricting the number of iterations to one (Hox 2002).
- 28 -
With respect to the fixed effects cluster-econometric techniques, we will employ the standard
within-estimator. However, another interesting estimator for fixed effects analysis has been
proposed by Pluemper and Troeger (2007). Their fixed effects vector decomposition estimator
(FEVD) is able to cope with endogeneity problems and neither has to eliminate all variables
which are invariant within clusters nor relies on instrumental variables as in the case of the
more established Hausman-Taylor approach. As study-level variables are by definition con-
stant within studies and instruments will hardly be available in the meta-analytical context,
FEVD here might be an appealing alternative to the standard within-estimator. FEVD estima-
tion is based on a three-step algorithm. In a first step, standard fixed effects estimates are
computed. Estimated cluster-specific effects then are regressed on within-invariant cluster-
level variables. The previously generated cluster-specific effects are thus decomposed into an
explained part – explained by the observed cluster-specific variables – and an unexplained
part. Subsequently, only the unexplained part, i.e. the step 2 residuals, is again included into
the original estimation equation. This equation, containing within-invariant, within-variant
variables and the unexplained part of the cluster-specific effect, now can be estimated by
POLS because the endogeneity problem has been resolved.22 However, as by construction the
FEVD estimator yields the same coefficient as standard within-estimation for all within-
variant model characteristics we only regard it as an option for the second stage regressions.
Notably, these first-stage regressions will be run twice. In a first step the semi-elasticities will
serve as the dependent variable and in a second step the t-values taken from the primary re-
gressions will be used as the dependent variable.23 This is in line with our major aim to fully
and transparently exploit all the information available from primary studies. The division by
standard errors guarantees that t-values are defined on a common dimensionless scale that is
independent of the original metric. Consequently, they are a valid methodological alternative
to the use of semi-elasticities. Using t-values as effect-size index is also advocated by Jarrell
and Stanley (1989, 2005). However, as has been underscored by Becker and Wu (2007), t-
values do not correspond to pure estimators of effect magnitude. In addition to effect magni-
tude, t-values are influenced by the precision of the estimate. They are thus estimators of ef-
fect significance rather than mere effect magnitude. Correspondingly, a positive partial corre- 22. The FEVD estimator has been applied in recent studies e.g. by Beckmann (2007) and Gatti (2008). 23. We will, however, not run WLS meta-regressions with t-values as the dependent variable. The use of t-
values as dependent variable can be interpreted as restricting the precision-weighting implied by WLS meta-regression to only the left-hand side of model (5). There should thus be no difference between estimation of model (5) with POLS or WLS when t-values are used as dependent variable since the WLS analytical weights by definition are equal to one. Heteroscedasticity now might, however, result from additional error components, i.e. unexplained heterogeneity as in model (6). We will therefore also employ the most flexible pooled estimator applying a weighting scheme: FGLS.
- 29 -
lation between t-values and a certain regressor would reveal that the respective study or model
characteristic on average fosters the identification of significant adverse tax effects on FDI.
6.2 Choosing the Preferred Meta-Regression Estimator
From the first-stage meta-regressions for both dependent variables only the preferred estima-
tors will form the basis for the subsequent general-to-specific analyses. Therefore we will –
after running a standard RESET test for misspecification – employ a set of suitable statistical
tests to systematically guide this choice.
In meta-analyses the conditional variance of the idiosyncratic error ( ),isV ε i isx z is directly
derived from the estimated variance matrix of the corresponding primary models. It is specific
to each model and depends mainly on factors such as sample size and estimation method. Ap-
parently, meta-analyses by construction have a problem with heteroscedasticity. The natural
starting point of our sequence of statistical tests is thus a standard test for heteroscedasticity
such as the White-test or the Breusch-Pagan test.
If the null of homoscedasticity is indeed rejected, this will in general favor estimation tech-
niques such as WLS estimates of models (5) and (6) or FGLS due to their explicit treatment
of the heteroscedasticity in the disturbances.24 Estimates of model (6) should be preferred
over those of model (5) if the Q-test for heterogeneity still rejects its null even after introduc-
tion of the moderator variables. However, none of these pooled estimators would be efficient
– and under certain conditions not even be consistent - in the case of observation dependence
due to unobserved study-level heterogeneity.
We therefore explicitly test for presence of unobserved cluster effects. However, we do not
solely argue based on the results of a standard Breusch-Pagan LM test. Instead, we in addition
adapt a similar test proposed by Wooldridge (2002) to the meta-analytic unbalanced cluster-
sample context. The test is robust to a non-normal distribution and – in particular – heterosce-
dasticity of the regression error terms. The test has power both against a random effects speci-
fication as well as against any kind of serial correlation in the errors. Offering these appealing
features the test seems predestined for application in the meta-analytical context.
Since under the null of no within-cluster correlation the variance matrix of the disturbances
for each study is diagonal, the Wooldridge-type test verifies if the average sum of the esti- 24. In order to estimate the weighting matrix supposed as unknown in FGLS we regress the squared POLS resi-
duals (in logs) on the full set of meta-regressors and then exponentiate the fitted values. This fairly flexible procedure has already been proposed by Harvey (1976). For details see Greene (2003).
- 30 -
mated non-diagonal elements across studies (scaled up by N ), 1
0,5
1 1 1
−−
= = = +∑∑ ∑
i iS SN
is iri s r s
N v v , is sta-
tistically different from zero. Therefore, we must scale it by its asymptotic standard error and
finally get the test statistic
1
1 1 10.521
1 1 1
ˆ ˆ
ˆ ˆ
−
= = = +
−
= = = +
⎡ ⎤⎛ ⎞⎢ ⎥⎜ ⎟⎢ ⎥⎝ ⎠⎣ ⎦
∑∑ ∑
∑ ∑ ∑
i i
i i
S SN
is iri s r s
S SN
is iri s r s
v v
v v
The subscript 1,...,=i N denotes the studies included in the analysis, while 1,...,= is S index-
es the observed estimates per study. isv represents the pooled OLS residuals used as estima-
tors for the pooled regression error isv . Under the null hypothesis this statistic is distributed
asymptotically as standard normal. Not rejecting the null clearly advocates the use of pooled
estimators. The test statistic’s null distribution, however, is only asymptotically valid. With
44=N (semi-elasticities) or 45=N (t-values) respectively we are clearly at the lower bound
of the sample size needed for reliable asymptotic inference. We are, however, convinced that
the test might give us additional interesting insights.
In a further step, we check whether there is an endogeneity problem linked to the unobserved
heterogeneity. If this is indeed the case, standard pooled or multilevel techniques will not
yield consistent estimates. Instead, fixed effects cluster-econometric approaches must be em-
ployed. However, the Hausman test as a standard tool for endogeneity checks should not be
applied in the meta-analytical context due to heteroscedasticity in the disturbances.
Instead, we use the robust artificial regression approach proposed in Wooldridge (2002)
which is equivalent to the Hausman test. The test is obtained by estimating an artificial re-
gression of the following form
= + +is isγ υis isw α z ς
where isγ is the quasi-demeaned version of isγ . ( )=is i isw x , z represents a vector containing
the quasi-demeaned study-level and model-level variables in ix and isz respectively. isz is
the fully demeaned version of all elements of isz which vary within studies.25 isυ is a homos-
cedastic and independent error term. The test rejects the null of no correlation between indi- 25. By definition all study-level variables in ix cannot vary within studies.
- 31 -
vidual error component and independent variables if the set of coefficients in vector ς is sta-
tistically significant.
Figure 2: Meta-analytical instruments and estimation strategy
Meta‐Study
Fixed‐effects MA
Random‐effects MA
Fixed‐effects MR
Mixed‐effects MR
WLSCluster/Panel Econometrics
WLS, FGLS,POLS
Random effects:REML, GLS
Fixed effects:FE; FEVD
Heteroscedasticity?BP Test
White Test
Misspecification?RESET test
BP LM Test;Test from
Wooldridge(2002)
Robust HausmanTest
Q‐Test
Q‐Test
Figure 2 illustrates the meta-analytical set of estimators and their conceptual relationships. Also
the tests applied to chose the most suitable among these estimators according to the meta-data at
hand is reflected in Figure 2. It thus draws on the explanations given in Section 5.1 and Section
6.2.
6.3 Expected Results
The insights sought for in this meta-study on the effect of taxation on FDI are twofold. With
the number of observations almost doubled it will be interesting to see whether the implica-
tions derived from primary empirical research today are the same as in the past. Due to the
sophisticated estimation strategy the meta-results will be of outstanding precision. On the oth-
er hand, we will employ the meta-analysis to learn about the influence of the newly included
meta-variables on the effect size estimates. With respect to the publication bias control we
might expect this phenomenon to be present. Publication bias has been detected in almost all
fields of empirical sciences and we do not expect FDI research to be an exemption.
As regards the more detailed differentiation of tax burden measures used in primary studies
we follow the reasoning put forward by Egger et al. (2008) and suggest that bilateral forward-
- 32 -
looking tax rates best capture the tax effect on transnational investments. Thus, we expect
studies relying on these bilateral measures to yield the highest effect size estimates, resulting
in a positive coefficient for the respective meta-regression dummy. As for the difference in
tax effect size estimates between aggregate and micro data studies we follow the line of
thought outlined in Becker et al. (2006) who argue that studies relying on aggregate FDI data
might suffer from simultaneity and endogeneity problems. If large unobservable parts of the
aggregate location conditions are not controlled for but put into the primary regression error,
this might bias measured tax effects. We are, however, not exactly sure about the sign of this
bias. In the same line of reasoning, controlling for country fixed effects in primary regression
models might indeed alter the estimated tax effects. As they should correct for possible biases
due to unobserved country effects, we expect the opposite sign as compared to the one found
for the “aggregate”-dummies. However, taxes are part of a country’s aggregate location cha-
racteristics and only limitedly time-variant. Consequently, their influence might be captured
to a large extent by the country-specific effects. Controlling for the latter might thus reduce
the significances of measured tax effects.
As to the inclusion of time dummies into primary regressions, its effects are more difficult to
anticipate. Time fixed effects generally absorb the influence that temporary (unobserved) ma-
croeconomic conditions exert on the considered dependent variable. However, if taxes do not
sufficiently vary across cross-section units (firms within countries or countries within consi-
dered regions) within the different time periods, then time fixed effects might mitigate identi-
fied tax effects or – precisely – their significance.
As to the inclusion of expenditure side variables into primary models, we derive our expecta-
tions about their effect on estimated tax impacts on FDI from some relatively straightforward
theoretical considerations. These argue within the scope of the Zodrow and Mieszkowski
(1986) framework and recent considerations by Benassy-Quéré et al. (2007) extending this
model.26 Benassy-Quéré et al. (2007) rely on two partial effects to compute the overall effect
of taxation on FDI. These partial effects measure the ceteris-paribus impacts of respectively
the company tax burden and public spending on FDI. Theoretically, they follow from the in-
ternational arbitrage equilibrium condition which requires equality of marginal after-tax re-
26. The framework builds on governments which maximize a representative agent’s utility U(x,G) with x denot-
ing the consumption of a (single) private good and G as public input used by both households and firms. G is financed by capital taxes and subject to the budget constraint. Production depends on capital, labor and public inputs. Marginal productivity is decreasing in all input factors. Inputs are complementary, but com-plementarity is decreasing with increasing inputs. For a comprehensive outline of the model see Benassy-Quéré et al. (2007) and for more fundamental version see also Zodrow and Mieszkowski (1986).
- 33 -
turn and real world market interest rate. Their estimated values are obtained from empirical
regressions including both the tax burden as well as expenditure side variables as regressors.
The overall effect of tax on FDI, however, according to the Zodrow and Mieszkowski frame-
work takes into account that tax and public expenditure variables are linked via the public
budget constraint. It therefore integrates the direct tax effect on FDI and the impact of increas-
ing public inputs financed by the tax receipts. Benassy-Quéré et al. (2007) show that the cete-
ris-paribus effects of taxes and of public spending still might combine to an overall effect if
the budget constraint is imposed ex post. And they indeed find that the computed overall ef-
fect of taxation on FDI is moderated by public spending, i.e. the direct adverse tax effect is
partially compensated for if tax receipts – at least partially – finance better public inputs. Sim-
ilar moderating effects can be expected for the inclusion of agglomeration variables.
Meta-analysis offers the opportunity to put the results reported by Benassy-Quéré et al. (2007)
into a broader context. Moreover, we do not have to compute the overall tax effect from par-
tial effects by imposing the budget constraint ex post. We instead regard tax effect estimates
from primary regressions which do not explicitly control for the (partial) influence of public
spending on FDI as estimates of the overall tax effect. This is in line with Brueckner (1979,
1982) who, however, examines the overall effect of public goods on property value in order
to test for the nonexistence of Tiebout equilibrium (also known as the Brueckner-Test). Bru-
eckner explicitly recognizes that the inclusion of both tax and expenditure side variables into
empirical estimation equations implicitly assumes their ceteris-paribus variation and thus neg-
lects the public budget constraint. He therefore solely estimates the public good coefficient
and excludes the tax variable from the regression. Thus, the tax rate is allowed to vary impli-
citly in response to a change of public spending policies in order to maintain budget balance
for the government. The coefficient for public good then gives the estimate for the respective
overall effect, covering any possible tax implications.27
Drawing an analogy to the context of this meta-study, we are able to statistically compare the
overall effect and the ceteris-paribus effect of taxation on FDI. Estimates for the first of both
effects are obtained from primary regressions without public spending control, i.e. allowing
for implicit variation of expenditures in response to a tax change. The second effect is instead 27. Since Brueckner (1979, 1982) argues within the scope of capitalization theory where local governments
maximize property values (and not agents’ utility) he can explicitly verify if equilibrium is reached by look-ing at the t-value of the public good variable coefficient estimate. In equilibrium, the expenditure side should have no significant effect on property value as the effect of a marginal increase in public goods is ex-actly compensated for by the implicit tax reaction. In the Zodrow and Mieszkowski world the moderating impact of public spending can be verified but not efficiency of public good provision. Here, capital is the left-hand variable. Utility and capital maxima, however, do not necessarily coincide.
- 34 -
obtained from regressions which simultaneously control for the tax burden and public expend-
itures and thus rest on the assumption of ceteris-paribus variation in these variables. By means
of a meta-regression dummy variable taking the value one if primary studies did not explicitly
control for public spending and zero otherwise we can statistically test for a moderating im-
pact of the expenditure side. In more econometric terms, we will see if the omission of ex-
penditure side variables biases the estimated tax effects. As we hypothesize public spending
to have a moderating effect, we expect a negative coefficient of the dummy variable, i.e. stu-
dies not controlling for the ceteris-paribus effect of public spending and thus letting it vary
implicitly with a tax change estimate lower adverse tax effects (in absolute terms).
7. Results
7.1 Classical Meta-Analysis
The results of the classical meta-analysis are shown in Figure 3. It contains the pooled effects
resulting from the random effects model only as the Q-test clearly indicates the presence of
considerable heterogeneity across studies (p < 0.01) for all four estimations. Thus, only the
random effects meta-analysis is viable as opposed to fixed effects estimation.
Figure 3: Results of classical random-effects meta-analysis
mprec1.358
(p < 0.01)
Abbreviations:min: pooled semi‐elasticity based on minimum estimates from each studymprec: pooled semi‐elasticity based on most precise estimates from each studymed: pooled semi‐elasticity based on median estimates from each studymax: pooled semi‐elasticity based on maximum estimates from each studyp represents p‐values from test for significance of pooled effects (H0: not significant)
med1.683
(p < 0.01)
min0.173
(p < 0.01)
max3.973
(p < 0.01)
We find that – independent of the percentile of semi-elasticity chosen (min, max or median) –
the effect of taxation on FDI is significantly different from zero (p < 0.01). The most precise
estimates are also highly significant. The pooled effect for the minimum semi-elasticities
amounts to 0.173. In contrast, the upper bound that is set by the pooled effect for the maxi-
mum effect sizes is 3.973. We calculate a pooled effect for the respective median effect sizes
of each study which amounts to 1.683. Selecting the most precise estimate from each study
gives a pooled semi-elasticity of 1.358. It should be noted that the results for minimum and
- 35 -
maximum estimates cannot merely be explained by outliers as these have been eliminated
from the sample before. Overall, the pooled estimates define a continuum of plausible values
as depicted in Figure 3.
7.2 Meta-Regressions
7.2.1 First Round Meta-Regressions
Semi-elasticities as dependent variable
Table 8 shows the results for the broad set of meta-regression estimators applied in a first
step. Columns (1) to (7) respectively represent the results from pooled OLS, feasible GLS,
fixed-effects meta-regression (FE WLS), mixed-effects meta-regression (ME WLS), fixed
effects cluster analysis, random effects cluster analysis via REML and random effects cluster
analysis based on GLS.
Clearly, many of the fundamental model characteristics controlled for in this first round esti-
mation show significant coefficients across different estimators. The employment of aggre-
gate data or aggregate data on affiliates, respectively, as opposed to micro level information
does indeed lead to higher primary tax-rate elasticities according to (nearly) all estimators. As
regards the influence of the tax rate chosen, the effective average tax rates (EATR) generally
yield higher primary estimates than those produced when country statutory tax rates are em-
ployed. This effect, however, is only significant in the case of pooled OLS, both fixed and
random effects meta-regression and random effects GLS. The evidence for the effective mar-
ginal tax rate (EMTR) is less clear-cut. Most coefficients indicate a negative but insignificant
impact on primary semi-elasticities, only fixed effects meta-regression and the within estima-
tor as well as random effects GLS show significant, but contradictory results.
36
Table 8: First round results for semi-elasticties
Dependent Variable: Semi-elasticities Pooled estimators Cluster-Econometric Estimators
Standard Pooled Traditional Meta-Regression Fixed
Effects Random Effects POLS FGLS FE WLS ME WLS FE RE (REML) RE (GLS) (1) (2) (3) (4) (5) (6) (7) Data Structure (Time Series) Cross-section 1.758 0.308 -0.979 -0.0164 2.710 0.424 0.615 (1.187) (0.908) (0.606) (0.936) (3.232) (1.419) (1.476) Panel Data -0.365 -0.385 -0.262 -0.778** -2.015 -0.928 -0.827 (0.350) (0.361) (0.518) (0.303) (1.569) (1.067) (1.119) Location Choice (Continuous Capital Data) -0.890 -0.692 0.052 -0.895 -5.284*** -1.574** -1.368** (1.059) (0.561) (0.369) (0.619) (1.186) (0.683) (0.615) Aggregation Level (Micro data) Aggregate Data 2.303*** 1.880*** 1.245*** 1.857*** - 1.489 1.628*** (0.698) (0.402) (0.417) (0.475) (0.923) (0.532) Aggregate Data on Affiliates 1.894** 1.393*** 0.922** 1.536*** 6.116*** 1.400* 1.408** (0.797) (0.451) (0.348) (0.519) (1.270) (0.829) (0.649) Tax Burden Measure (Country Statutory Rate) Micro Average Tax Rate -0.360 -0.227 1.875*** 0.170 0.151 -0.057 -0.091 (1.003) (0.628) (0.664) (1.030) (1.067) (0.647) (0.675) Macro Average Tax Rate -1.094 -1.380*** 1.074** -1.722** -1.668 -1.693 -1.596 (0.667) (0.399) (0.419) (0.657) (1.697) (1.140) (1.073) State Statutory Tax Rate 2.615** 2.705** 3.724*** 2.553*** - 2.773*** 2.717** (1.064) (1.274) (0.846) (0.934) (1.061) (1.151) Marginal Effective Tax Rate -0.488 -0.671 1.728*** -0.337 -0.868* -0.755 -0.726* (0.613) (0.473) (0.374) (0.539) (0.444) (0.519) (0.379) Average Effective Tax Rate 1.057** 0.508 1.611** 1.101** 0.335 0.656 0.727** (0.503) (0.360) (0.608) (0.531) (0.411) (0.556) (0.355) Constant 0.942 1.625** -0.812 1.579** 4.012*** 2.277 2.051 (0.939) (0.659) (0.883) (0.689) (1.484) (1.453) (1.258) Observations 708 708 707 707 708 708 708 R2 0.145 0.179 0.281 0.134 0.047 - - F-/Wald-test (H0: all coefficients = 0) p = 0.000 p = 0.000 p = 0.000 p = 0.000 p = 0.000 p = 0.000 p = 0.001 Clustered standard errors (POLS, FGLS, WLS) or robust standard errors (FE, REML, RE GLS) in parentheses. ***/**/* denotes significance at the 1%/5% /10% level. Primary semi-elasticities haven been multiplied by (-1). All study/model characteristics are coded as dummy variables. Thus, a base model represents the characteristics redundant to the variables explicitly included. The base characteristics are indicated in parenthesis for each study dimension. Estimated coefficients of the dummies indicate the effect on primary semi-elasticities of choosing a characteristic in lieu of the base specification. Abbreviations for the employed estimators are explained in the text (top of previous page).
37
As regards the influence of macroeconomic average tax rates as compared to country statutory
rates, most coefficient estimates are negative, but only sometimes significant. Fixed effects
meta-regression again shows significant but opposite results for this characteristic. Employing
microeconomic average tax rates does not yield different primary semi-elasticities as com-
pared to statutory tax rates according to all meta-regression estimators, again with the excep-
tion of traditional fixed effects meta-regression. State statutory tax rates lead to higher tax
effect estimates as compared to the use of tax rates at the federal level. Whether cross-
sectional or panel data is used as opposed to time series analysis is not relevant according to
most estimators. Precisely, only mixed effects meta-regression indicates a significant negative
impact of using panel data. Studies which examine the location choice of companies syste-
matically measure lower tax effects. For all cluster econometric approaches this effect is sig-
nificant. Accordingly, the localization of economic rents as compared to marginal investment
decisions seems less influenced by taxes. This result would be in line with the findings of De
Mooij and Ederveen (2006). It will, however, need further investigation to be clearly verified.
Table 9: Choosing the preferred meta-regression estimator (semi-elasticities)
Data characteristic Test Result Preferred estimators
1. Heteroscedasticity? Breusch Pagan test χ²(1)= 6.98; p < 0.01
FGLS, FE WLS, ME WLS
White test χ²(30)= 121.57; p < 0.01
2. Unexplained
heterogeneity? Q - test Q = 6726.37; p < 0.01 FGLS, ME WLS
3. Study-specific ef-
fects?
Breusch Pagan LM test χ²(1)= 12.63; p < 0.01 Cluster-econometrics (FE, RE)
Wooldridge-type test SN = 1.86; p = 0.03
4. Endogeneity? Robust Hausman test F(8, 41) = 19.71; p <0.01 FE
1. Heteroscedasticity: Breusch Pagan and White test are both based on test statistics which are χ²-distributed under the null of no heteroscedasticity. In the present case both tests reject the null of no heteroscedasticity at very high significance levels. 2. Unexplained heterogeneity: The Q-test is based on a test statistic which is χ²-distributed under the null of no heterogeneity. Here, the null of no heterogeneity is rejected at very high signi-ficance levels. 3. Study-specific effects: The Breusch Pagan LM and the Wooldridge-type assume a null of no unobserved study-specific effects. Under the null the Breusch Pagan LM test is χ²-distributed, while the Wool-dridge-type test is asymptotically distributed as standard normal under the null. Both tests reject the null at high significance levels. 4. Endogeneity: F-test in the context of an artifical regression following Wooldridge (2002), also see the explanations in the text. The null of no correlation between the study-specific unobservable effect and the moderator variables is rejected at high significance levels.
Although most of the applied meta-regression estimators give congruent indications as re-
gards the influence of study characteristics on primary tax effect estimates, they rely on dif-
ferent assumptions regarding the characteristics of the meta-data, as has been outlined in sec-
38
tion 5. We therefore employ the series of tests described in section 6 in order to find the esti-
mator which is most suitable for the actual data at hand. The test results are given in Table 9.
First, we run a standard RESET test on the basic specification employment in these first round
estimations. The result ( F(3,694) 13.93= , p < 0.01) forces us to reject the null of correct spe-
cification. We truncate our meta-sample at the extremes and successively rerun the test until
the null is no longer rejected. Comparing the implications of running the regressions on this
truncated sample, no qualitative differences occur. We thus go on with the meta-regression
based on the full sample. In a next step we find clear evidence for heteroscedasticity in the
meta-regression error terms by means of a Breusch Pagan LM test ( 2χ (1) = 6.98, p < 0.01)
and a White test ( 2χ (30) = 121.57, p < 0.01). This result clearly favors meta-regression esti-
mators which explicitly cope with heteroscedasticity. As a Q-test (Q = 6726.37, p < 0.01) run
after the mixed-effects meta-regression identifies remaining unexplained heterogeneity, we
can exclude fixed effects WLS from the set of applicable estimators. In order to assess wheth-
er meta-regression models should account for a nested error structure, we add to the standard
Breusch Pagan LM test ( 2χ (1) = 12.63, p < 0.01) the Wooldridge-type test (test = 1.86, p =
0.03) outlined above. Apparently, unobserved study-effects are present. Cluster econometric
techniques are thus preferred over pooled estimators. As we find clear evidence for endogene-
ity, only the fixed effects cluster estimator is viable. To gain insights into the effects of va-
riables which are constant within studies we also apply the FEVD technique. Another robust-
ness check includes a traditional mixed effects meta-regression based on WLS.
T-values as dependent variable
Replacing semi-elasticities by primary t-values on the left-hand side of the meta-regression
equation implies that resulting meta-coefficients do not only measure the impact of consi-
dered meta-regressors on mere effect size. In contrast, t-values by definition also contain in-
formation about primary estimate precision. Thus, both types of analysis should be regarded
as complementary rather than substitutes. In the light of the insights gained from both analys-
es some statements made solely on the grounds of semi-elasticities might be put into perspec-
tive.
39
Table 10: First round results for t-values
Dependent Variable: T-values Pooled Estimators Cluster-Econometric Estimators
Fixed Effects Random Effects POLS FGLS FE RE (REML) RE (GLS) (1) (2) (3) (4) (5) Data Structure (Time Se-ries) Cross-section 2.012** 2.147*** 0.697 -0.007 0.006 (0.895) (0.641) (1.193) (1.271) (1.021) Panel Data 1.078** 0.920* 1.024 0.229 0.228 (0.454) (0.484) (0.752) (0.896) (0.731) Location Choice (Continuous Capital Data) -0.839 -0.630 -1.348*** -0.901* -0.897** (0.523) (0.382) (0.402) (0.484) (0.351) Aggregation Level (Micro data) Aggregate Data -0.131 -0.127 - -0.233 -0.241 (0.657) (0.681) (0.881) (0.735) Aggregate Data on Affiliates -1.051** -1.247*** 1.046** -0.681 -0.698 (0.413) (0.397) (0.445) (0.768) (0.556) Tax Burden Measure (Country Statutory Rate) Micro Average Tax Rate -0.832 -1.088* 0.883 0.565 0.557 (1.252) (0.570) (0.697) (0.498) (0.619) Macro Average Tax Rate -1.374 -1.585** -2.369*** -1.688* -1.681** (1.127) (0.648) (0.750) (0.862) (0.729) State Statutory Tax Rate -2.042** -2.030*** - -0.914 -0.922 (0.892) (0.546) (1.162) (0.780) Marginal Effective Tax Rate -0.994 -1.077* -0.157 -0.237 -0.241 (1.147) (0.603) (0.296) (0.356) (0.302) Average Effective Tax Rate 0.0862 0.261 0.230 0.204 0.204 (0.972) (0.399) (0.374) (0.393) (0.343) Constant 2.809** 2.952*** 1.720** 3.111** 3.119*** (1.139) (0.942) (0.708) (1.232) (0.928) Observations 729 729 729 729 729 R2 0.121 0.190 0.022 F-/Wald-test (H0: all coefficients = 0) p = 0.000 p = 0.000 p = 0.001 p = 0.089 p = 0.003 Clustered standard errors (POLS, FGLS) or robust standard errors (FE, REML, RE GLS) in parentheses. ***/**/* denotes signi-ficance at the 1%/5% /10% level. Primary t-values have been multiplied by (-1). All study/model characteristics are coded as dummy variables. Thus, a base model represents the characteristics redundant to the variables explicitly included. The base cha-racteristics are indicated in parenthesis for each study dimension. Estimated coefficients of the dummies indicate the effect on primary t-values of choosing a characteristic in lieu of the base specification. Abbreviations for the employed estimators are ex-plained in the text.
The first round results for the meta-regressions which use primary t- and z-values as depen-
dent variable are given in Table 10. As compared to the first round results for the semi-elasti-
cities the influence of many fundamental primary characteristics are less significant. Neither
the employment of effective tax rates (as opposed to the country statutory rate) nor the use of
aggregate data (as opposed to micro data) has a significant impact on t-values according to all
or all but one of the applied estimators. Strikingly, the impact of model and study characteris-
tics on semi-elasticities or t-values respectively is not always of the same sign. The state tax
rate here shows negative coefficients and the effect is significant. According to most estima-
tors, the macroeconomic average tax rate is a rather crude measure to capture tax incentives. It
leads to much lower primary t-values as compared to the statutory rate. Furthermore, aggre-
40
gate studies do not yield higher t-values as compared to micro data studies. As for studies
using aggregate data on affiliates the evidence is not quite clear, since we find significant ef-
fects with different signs according to the different estimators. However, in any case the con-
siderable bias in semi-elasticities does not seem to fully translate into higher t-values, i.e. sig-
nificance of the tax effects. The presumed upward bias due to endogeneity problems in aggre-
gate studies thus seems to be counteracted by higher standard errors of the primary tax effect
estimates. Attributing this observation to problems of measurement error in the context of
aggregate studies seems plausible. Again, the influence of taxes is less important for the deci-
sion on the location of economic rents as compared to investment decisions conditional on
location choice. So we measure generally lower t-values for studies analyzing location choice
as compared to studies examining continuous FDI. However, similar to what we find in the
analysis focusing on semi-elasticities, the significance of this effect becomes not absolutely
clear by looking at the results across Table 10.
Table 11: Choosing the preferred meta-regression estimator (t-values)
Data characteristic Test Result Preferred estimators
1. Heteroscedasticity? Breusch Pagan LM test χ²(1)= 10.19, p < 0.01
FGLS White test χ²(30)= 41.18, p = 0.08
2. Study-specific ef-
fects?
Breusch Pagan LM test χ²(1)= 102.95, p < 0.01 Cluster-econometrics (FE, RE)
Wooldridge-type test test = 1.54, p = 0.06
3. Endogeneity? Robust Hausman test F(8, 41) = 3.17, p < 0.01 FE
See Footnote of Table 9 for explanations.
Starting with the series of test, we compute a RESET test for misspecification
( F(3,715) 2.43= , p =0.06) and find only relatively weak evidence for a rejection of the null.
However, we still repeat the same procedure as in the case of semi-elasticities until the null is
clearly accepted. Again, we find no qualitative differences in results for the regressions on the
truncated sample and go on with the analysis on the full meta-dataset. The subsequent tests
and resulting statistics used to identify the preferred meta-regression estimator for the analysis
of t-values are indicated in Table 11. Again, we find clear evidence for heteroscedasticity.
Study-specific unobserved effects are present according to the standard Breusch Pagan test as
well as to the more robust Wooldridge-type test. The robust Hausman test based on the artifi-
cial regression approach, however, reveals an endogeneity problem linked to the unobser-
41
vables. Thus, our preferred estimator is the fixed effects cluster estimator. In order to gain
insights on the effect of within-invariant variables we again apply the FEVD technique.
Another robustness check will also look at FGLS results.
7.2.2 Second Round Meta-Regressions
Semi-elasticities as dependent variable
The results for the second round meta-regression with semi-elasticities on the left-hand side
are indicated in Table 12. First, we extend the first round specification by including the newly
coded model characteristics (Column 1). Obviously, splitting the effective forward-looking
tax rates in bilateral and unilateral tax burden measures reveals that particularly the bilateral
measures seem to capture adverse tax effects on foreign direct investments. Both the bilateral
marginal tax burden measures and the bilateral average tax rate significantly raise primary tax
effect estimates as compared to statutory rates. Strikingly, employing the domestic unilateral
effective tax rates (EATR and EMTR) leads to significantly lower estimated semi-elasticities
as compared to the use of statutory rates. These measures might thus indeed mix aspects of
competitive (tax) advantage for multinationals over domestic firms and pure adverse tax ef-
fects. This supports the reasoning of Egger et al. (2008), who explicitly prefer bilateral tax
burden measures in the FDI context. However, the significantly negative coefficient of the
effective average unilateral tax rates is not robust across specifications. This seems intuitively
plausible since the EATR is to a considerable extent driven by the statutory tax rate.
Country fixed effects significantly reduce primary tax effect estimates. They thus do indeed
show the opposite sign of the coefficient we find for aggregate studies. The latter is still high-
ly significant and positive for studies based on aggregate data on affiliates. In the left four
columns of Table 12 we could not estimate a coefficient for aggregate studies due to a lack of
within-study variation. The inclusion of time fixed effects, however, does exert no effect on
estimated semi-elasticities. Strikingly, regarding the implications of considering the public
expenditure side in primary regressions, we do not identify a significant bias. Also the sign of
the effect of not-including a public spending control is not as expected. In other words, allow-
ing public spending to adjust to tax changes and thus not including it into the primary regres-
sion does not lower the estimated adverse tax effects. As this result contradicts theoretical
presumptions, it might at least in parts be explained by hitherto insufficiently precise proxies
for relevant public inputs from the companies’ point of view.
42
Table 12: Second round results for semi-elasticities
Dependent Variable: Semi-elasticities Extensions Robustness Checks
FE-Extension 1
FE-Extension 2
FE-Extension 3
FE-Extension 4
Robustness Check 1
Robustness Check 2
Robustness Check 3
(1) (2) (3) (4) (5) (6) (7) Data Structure (Time Series) Cross-section 2.263 2.285 3.953 5.167 5.167 -1.087 -1.805 (3.233) (3.219) (2.523) (3.585) (3.659) (0.788) (2.113) Panel Data -0.713 -0.368 0.746 1.153 1.153 0.155 -2.594 (1.601) (1.730) (1.316) (2.269) (1.268) (0.752) (1.776) Location Choice (Continuous Capital Data) -5.284*** -5.067*** -3.515*** -4.612*** -4.612*** -4.612*** -0.968 (1.189) (1.256) (1.233) (1.282) (0.922) (0.953) (0.662) Aggregation Level (Micro data) Aggregate Data - - - - 3.945*** 2.367*** 2.515*** (0.131) (0.183) (0.574) Aggregate Data on Affiliates 5.548*** 5.304*** 3.673*** 4.861*** 4.861*** 4.861*** 2.347*** (1.243) (1.327) (1.282) (1.347) (1.202) (1.280) (0.474) Tax Burden Measure (Country Statutory Rate) Micro Average Tax Rate -0.237 -0.203 -0.00450 -0.0478 -0.0478 -0.0478 -0.130 (1.074) (1.065) (0.735) (1.069) (0.578) (0.593) (0.698) Macro Average Tax Rate -2.594 -2.725* -0.960 -2.553 -2.553 -2.247*** -0.522 (1.627) (1.651) (0.943) (1.659) (1.561) (0.753) (1.266) State Statutory Tax Rate - - - - 3.495*** 4.467*** 3.845* (0.289) (0.775) (2.065) Marginal Effective Tax Rate -1.424** -1.411** -0.851** -1.405** -1.405*** -1.405*** -0.235 (0.572) (0.569) (0.348) (0.570) (0.396) (0.425) (0.665) Average Effective Tax Rate -0.666 -0.655 -0.733** -0.663 -0.663* -0.663* 0.086 (0.438) (0.433) (0.319) (0.426) (0.397) (0.390) (0.343) Bilateral Marginal Effective Tax Rate 1.099*** 1.123*** 0.949*** 0.979*** 0.979*** 0.979*** 0.377 (0.315) (0.315) (0.257) (0.328) (0.309) (0.300) (0.717) Bilateral Average Effective Tax Rate 3.566*** 3.562*** 3.144*** 3.465*** 3.465*** 3.465*** 1.806*** (0.538) (0.538) (0.479) (0.544) (0.438) (0.452) (0.604) No Control for Public Spending (Control) 0.460 0.466 0.253 0.507 0.507 0.507 0.408 (0.863) (0.864) (0.533) (0.942) (0.766) (0.682) (0.373)
43
Dependent Variable: Semi-elasticities Extensions Robustness Checks
FE-Extension 1
FE-Extension 2
FE-Extension 3
FE-Extension 4
Robustness Check 1
Robustness Check 2
Robustness Check 3
(1) (2) (3) (4) (5) (6) (7) Specification (no Fixed Effects) Time Fixed Effects -0.332 -0.422 0.389 -0.381 -0.381 -0.381 -0.970* (0.530) (0.512) (0.466) (0.513) (0.497) (1.014) (0.518) Country Fixed Effects -1.302*** -1.312*** -1.062*** -1.246*** -1.246*** -1.246*** -0.468 (0.298) (0.297) (0.239) (0.325) (0.246) (0.440) (0.311) OLS Estimation (other than OLS) -1.229*** -1.229*** -1.229*** -0.655* (0.442) (0.386) (0.399) (0.339) Type of Industry (not specified) Manufacturing Industry 0.971* 0.971** 0.971** 0.856** (0.515) (0.452) (0.478) (0.405) Financial Services 0.891 0.891* 0.891** 1.599* (0.705) (0.465) (0.411) (0.900) Finance (not specified) Retained earnings 1.168 1.168 1.168 0.083 (1.437) (0.949) (1.101) (0.843) Transfers of Funds 0.667 0.667 0.667 -0.453 (1.210) (0.913) (1.060) (0.520) Double Taxation Relief System (not specified) Credit System 0.211 0.211 0.211 -0.145 (0.788) (0.496) (0.499) (0.615) Exemption System 0.107 0.107 0.107 -0.498 (0.787) (0.496) (0.505) (0.521) Control Variables (not included) Control for Home Tax -1.979 -1.979* -1.979* -0.221 (1.829) (1.089) (1.171) (0.496) GDP -0.197 -0.197 -1.016** 0.178 (1.943) (2.300) (0.413) (0.585) Openness -1.388*** -1.388*** -2.004*** -1.822*** (0.353) (0.373) (0.392) (0.224) Agglomeration Effects 1.013 1.013 -0.001 0.053 (2.160) (1.568) (0.214) (0.541) Exchange Rate 1.178 1.178 1.178 0.695 (1.808) (2.096) (1.586) (0.765) Wage -1.258 -1.258 0.200 0.456 (0.889) (0.926) (0.511) (0.419)
44
Dependent Variable: Semi-elasticities Extensions Robustness Checks
FE-Extension 1
FE-Extension 2
FE-Extension 3
FE-Extension 4
Robustness Check 1
Robustness Check 2
Robustness Check 3
(1) (2) (3) (4) (5) (6) (7) FDI Target Regions (no specific region) Europe - 2.743*** 0.963*** 0.080 (0.178) (0.184) (0.423) United States - -1.175*** -3.253*** -3.242* (0.192) (0.592) (1.858) Average Sample Year -0.105 -0.105 -0.105 0.0258 (0.137) (0.0911) (0.116) (0.0563) Publication Bias Standard Error 0.0709 0.0957 0.0957 0.0957 0.498*** (0.123) (0.113) (0.0754) (0.0753) (0.139) Standard Error * D_pos 0.578*** (0.109) Standard Error * D_neg -1.284*** (0.147) Constant 3.865** 3.516** 1.557 212.3 209.1*** 213.3*** -47.92 (1.647) (1.772) (1.303) (271.9) (0.133) (0.708) (111.4) Observations 700 699 699 699 699 699 699 R2 0.075 0.076 0.527 0.099 0.315 0.315 0.284 F-test (H0: all coefficients = 0) p = 0.000 p = 0.000 p = 0.000 p = 0.000 p = 0.000 p = 0.000 p = 0.000 Robust standard errors in parentheses. ***/**/* denotes significance at the 1%/5% /10% level. Primary semi-elasticities haven been multiplied by (-1). All extensions are esti-mated by fixed effects (within) estimation. Robustness checks 1 and 2 employ the fixed effects vector decomposition technique (FEVD), see Pluemper and Troeger (2007). Robustness check 1 includes only within study invariant characteristics into the 2nd stage of the FEVD estimator, robustness check 2 uses FEVD with also the least within-variant variables included into the 2nd stage. All study/model characteristics (except average sample year and standard error) are coded as dummy variables. Thus, a base model represents the characteristics redundant to the variables explicitly included. The base characteristics are indicated in parenthesis for each study dimension. Estimated coefficients of the dummies indicate the effect on primary semi-elasticities of choosing a characteristic in lieu of the base specification.
45
Another aspect which has so far not been considered in surveys of the literature on FDI and
taxation is publication bias. In Column 2 we consider unidirectional publication bias by in-
cluding the standard error of the primary estimates into the meta-regression. However, studies
on FDI and taxation do not seem to be significantly biased according to this approach. Re-
searchers do not try to raise effect size estimates in response to imprecise estimation tech-
niques. The conclusions are, however, quite different if we look at what is found for the more
sophisticated approach introduced by Bom and Ligthart (2008). Positive tax effect estimates
apparently are pushed even more if estimates tend to be imprecise. Negative tax effect esti-
mates are even more negative in response to high standard errors. Both effects are significant.
Bidirectional publication bias thus seems to be present. However, it is not persuasive that re-
searchers who identify attracting effects of taxation on FDI tend to insist on their findings and
try to push them towards significance. According to the 2R of the regression in Column 3 the
approach by Bom and Ligthart leads to a veritable jump in the explained variation of primary
estimates. However, by construction, the interaction terms (Standard Error * D_pos) and
(Standard Error * D_neg) show a fairly high (within-) variation as compared to the other re-
gressors. This, however, fosters significance of the two examined correlations. For this reason
we hesitate to interpret the corresponding results as revealing causality and rely on the unidi-
rectional approach in the remaining meta-regressions.
In Column 4 of Table 12 we estimate a fully extended meta-regression with additional study
and model characteristics which might admittedly have an impact on primary study results
(De Mooij and Ederveen 2003, 2005, 2006). Almost all results are in line with the findings of
De Mooij and Ederveen (2003, 2005, 2006). The system of double taxation avoidance in the
parent country or the mode of finance (retained earnings vs. transfer of funds) does not influ-
ence the estimated tax effects. Primary studies controlling for the openness of the target econ-
omies yield significantly smaller semi-elasticities while most of other primary control va-
riables (GDP, Agglomeration effects, Exchange rates and wage) do not have significant im-
pacts on measured tax effects.28 The mean sample year also has no impact on primary esti-
mates of the tax effect on FDI. Controlling for home country taxes exerts no significant effect
on primary tax effect estimates. The respective meta-coefficient, however, has a negative
sign. Interestingly, studies who apply OLS as opposed to any other estimation technique (e.g.
instrumental variable estimators) significantly measure lower tax effects on FDI. Finally, we
find that studies explicitly concentrating on manufacturing industries find higher effect esti- 28. De Mooij and Ederveen (2005), however, found that controlling for agglomeration effects had significant
impacts on primary estimates.
46
mates as opposed to those which do not specify the sectors covered by their analyses. The
inclusion of all these meta-regressors do not change the results found before. Please note that
all our second round results show no significant effect of the data structure – panel, cross-
section or time series – on primary results. This is in contrast to earlier findings of De Mooij
and Ederveen and might be due to the partly modified study classification.
The last three columns of Table 12 represent robustness checks. In Column 5 we estimate the
full model via FEVD. This is particularly interesting as it allows for the identification of the
effects of study-level variables (or within-invariant model-level characteristics) on primary
semi-elasticities. Besides the statutory tax rate the focus, of course, is on the coefficient esti-
mate for aggregate data studies. For both of these variables we find the first stage regression
results confirmed as the measured effects are positive and highly significant. In the course of
this robustness check we also take a look at FDI target regions considered in primary work.
Precisely, we test whether tax effects significantly differ if studies focus exclusively either on
European countries or the US as FDI target. This was not possible as long as standard fixed
effects estimation was employed. The benchmark model for both characteristics is defined by
those studies which examine a mix of regions, typically a broad set of OECD countries. Ap-
parently estimated adverse tax effects on FDI are stronger for only Europe as compared to the
benchmark, while the tax effect measured for the US is below the benchmark. As to the im-
plications regarding the estimated coefficients of the other meta-regressors, only changes in
significance might occur. This is the case for the effects estimated for the average effective
tax rate and for models estimated solely on information about financial service industries. In
Column 6 we make use of the possibility to include not only purely within-invariant variables
into the second stage of the FEVD estimator, but we further add the most rarely changing
model-level variables to this stage of the estimation procedure. This, however, does not
change the results except that the effect for GDP control now turns out to be significant. In
Column 7 we again look at results from WLS mixed effects meta-regression. We find that
significance for some of the effects identified is lost. Particularly, this holds for the influence
of target regions on primary semi-elasticities as well as for the effects of employing bilateral
marginal tax rates or location choice analysis. Additionally, we now find the effects of the
inclusion of time fixed effects into primary regressions to be significant, while fixed effects
no longer seem to change primary results significantly. But many results also do not change
with respect to significance. This is particularly the case for the effect of using aggregate data
and bilateral average tax rates.
47
T-values as dependent variable
For the second round computations with t-values on the left-hand side (Results: Table 13) a
set of specifications and estimators similar to the one applied to semi-elasticities is chosen.
An important difference, however, is that we cannot control for the (non-)existence of publi-
cation bias in the literature by including primary standard errors on the right-hand side of the
meta-regression because this would imply simultaneity problems. As a solution we simply
include a dummy reflecting the publication status of primary work into the regressions. As
publication status represents a study-level characteristic which is by definition invariant with-
in studies we unfortunately cannot use standard within estimation to examine publication bias.
Therefore we address this issue later in the course of two robustness checks we perform on
the basis of FEVD (see Columns (3) and (4) in Table 13). As a third robustness check Column
(5) in Table 13 is estimated by pooled FGLS. Generally, the results – as displayed in Table 13
– are again fairly robust across specifications. However, for some regressors we see switches
with respect to the apparent significance of the effects they exert on t-values.
In principle, the presumptions from the first round estimates are confirmed. While aggregate
studies produce higher semi-elasticities as compared to micro data studies they do not yield
more significant tax effects. According to the results in Columns (3) and (4), t-values are even
significantly mitigated. The picture is somewhat different in the case of studies using aggre-
gate affiliate data. Here, a higher significance of primary tax effects is generally identified. To
investigate this point in more depth we run a simple POLS regression of primary effect stan-
dard errors on the examined study and model characteristics.29 Indeed we find that aggregate
studies as well as studies based on aggregate affiliate data produce significantly higher stan-
dard errors as compared to micro data studies. And indeed, this effect is most pronounced for
aggregate studies.
According to all specifications estimated by the consistent fixed effects estimators, we finally
confirm that the decision where to locate economic rents indeed responds less to taxes than
continuous investment decisions at the margin. This proposition has already been put forward
by De Mooij and Ederveen (2006). It, however, contradicts the general presumptions drawn
from the fact that estimated coefficient for marginal tax burden measures (e.g. EMTR) are
generally lower (again in absolute terms) than those for their inframarginal ones. This often is
seen as an implicit indication for relatively low tax responsiveness of marginal investment
decisions (see. e.g. Devereux and Lockwood 2006). A plausible explanation of the results 29. Results are available upon request.
48
obtained from the meta-regressions might be based on the fact that decisions at the margin are
by definition very difficult to capture – even on the basis of high quality continuous capital
data. For example, Overesch and Wamser (2008b) found that the EATR effect on continuous
investment decisions conditional on location choice is still higher than the EMTR effect.
Table 13: Second round results for t-values
Dependent Variable: T-values Extensions Robustness Checks
FE-Extension 1
FE-Extension 2
Robustness Check 1
Robustness Check 2
Robustness Check 3
(1) (2) (3) (4) (5) Data Structure (Time Series) Cross-section -1.311 -0.275 -0.275 -3.246*** -2.582 (1.300) (1.622) (1.427) (0.472) (1.631) Panel Data 1.009 0.876 0.876 -1.881*** -1.781 (1.167) (1.333) (0.724) (0.465) (1.359) Location Choice (Continuous Capital Data) -1.554*** -1.697*** -1.697*** -1.697*** -1.156** (0.466) (0.542) (0.352) (0.347) (0.492) Aggregation Level (Micro data) Aggregate Data -0.273* -0.428*** 0.826 (0.154) (0.113) (0.628) Aggregate Data on Affiliates 1.084 1.276* 1.276*** 1.276*** 0.813* (0.668) (0.731) (0.473) (0.465) (0.459) Tax Burden Measure (Country Statutory Rate) Micro Average Tax Rate 0.537 0.657 0.657 0.657 -0.814 (0.642) (0.591) (0.443) (0.461) (0.519) Macro Average Tax Rate -4.002*** -4.025*** -4.025*** -1.799*** -1.369* (0.607) (0.714) (0.915) (0.614) (0.691) State Statutory Tax Rate - - -1.428*** -0.460 1.373 (0.227) (0.326) (1.278) Marginal Effective Tax Rate -0.779*** -0.646** -0.646** -0.646** -0.989** (0.279) (0.298) (0.271) (0.265) (0.470) Average Effective Tax Rate -0.698* -0.548 -0.548 -0.548 -0.0743 (0.364) (0.356) (0.442) (0.414) (0.290) Bilateral Marginal Effective Tax Rate 2.207*** 2.248*** 2.248*** 2.248*** 0.718* (0.369) (0.348) (0.342) (0.294) (0.373) Bilateral Average Effective Tax Rate 3.692*** 3.570*** 3.570*** 3.570*** 0.919 (0.542) (0.521) (0.476) (0.431) (0.555) No Control for Public Spending (Con-trol) 0.506 0.0810 0.0810 0.0810 -0.148 (0.519) (0.316) (0.247) (0.452) (0.322) Specification (no Fixed Effects) Time Fixed Effects -1.882*** -2.211** -2.211* -2.211** -1.066 (0.575) (0.877) (1.166) (0.893) (0.652) Country Fixed Effects 0.0148 0.147 0.147 0.147 -0.755*** (0.891) (0.921) (0.497) (0.491) (0.248) OLS Estimation (other than OLS) -0.749** -0.749*** -0.749*** -0.216 (0.298) (0.162) (0.156) (0.330) Type of Industry (not specified) Manufacturing Industry -0.483 -0.483 -0.483 0.576 (0.397) (0.393) (0.334) (0.387) Financial Services 1.148** 1.148** 1.148*** 0.355 (0.514) (0.463) (0.341) (1.005) Finance (not specified) Retained earnings 0.176 0.176 0.176 0.0372 (0.580) (0.719) (0.744) (0.680) Transfers of Funds -0.218 -0.218 -0.218 -0.926 (0.481) (0.621) (0.641) (0.622)
49
Dependent Variable: T-values Extensions Robustness Checks
FE-Extension 1
FE-Extension 2
Robustness Check 1
Robustness Check 2
Robustness Check 3
(1) (2) (3) (4) (5) Double Taxation Relief System (not specified) Credit System -0.125 -0.125 -0.125 -0.116 (0.460) (0.381) (0.336) (0.467) Exemption System -0.605 -0.605* -0.605* -1.168** (0.456) (0.357) (0.309) (0.460) Control Variables (not included) Control for Home Tax -0.759 -0.759 -0.759 -1.334** (0.523) (0.741) (0.733) (0.564) GDP 0.972 0.972 0.647*** -0.516 (1.208) (0.772) (0.249) (0.558) Openness -2.038*** -2.038*** -1.335*** -1.235*** (0.407) (0.531) (0.241) (0.416) Agglomeration Effects -1.755 -1.755 0.676* -0.621* (2.378) (2.471) (0.362) (0.327) Exchange Rate -0.223 -0.223 -0.223 1.667* (1.319) (1.088) (1.256) (0.990) Wage 1.806 1.806* 0.826*** 0.181 (1.329) (0.976) (0.312) (0.416) FDI Target Regions (no specific region) Europe 0.159 0.155 -0.520 (0.0988) (0.121) (0.419) United States 0.162 -2.166*** -3.459*** (0.138) (0.326) (1.140) Average Sample Year -0.0127 -0.0127 -0.0127 0.0722 (0.0568) (0.0692) (0.0725) (0.0492) Publication Bias Unpublished (Published) - -1.186*** -0.713*** -0.936* (0.219) (0.155) (0.477) Constant 2.714*** 28.90 29.56*** 32.19*** -135.9 (0.923) (112.9) (0.257) (0.474) (96.74) Observations 721 721 721 721 721 R2 0.086 0.130 0.456 0.456 0.313 F-test (H0: all coefficients = 0) p = 0.000 p = 0.000 p = 0.000 p = 0.000 p = 0.000 Robust standard errors in parentheses. ***/**/* denotes significance at the 1%/5% /10% level. Primary t-values have been mul-tiplied by (-1). All extensions are estimated by fixed effects (within) estimation. Robustness checks 1 and 2 employ the fixed effects vector decomposition technique (FEVD), see Pluemper and Troeger (2007). Robustness check 1 includes only within study invariant characteristics into the 2nd stage of the FEVD estimator, robustness check 2 uses FEVD with also the least within-variant variables included into the 2nd stage. Robustness check 3 is estimated by pooled FGLS. All study/model characteristics (except average sample year) are coded as dummy variables. Thus, a base model represents the characteristics redundant to the variables explicitly included. The base characteristics are indicated in parenthesis for each study dimension. Estimated coeffi-cients of the dummies indicate the effect on primary t-values of choosing a characteristic in lieu of the base specification.
With respect to the influence of the data structure on primary results, the results in Column (4)
suggest that both panel and cross-section significantly reduce primary t-values as compared to
time-series studies. However, the majority of results do not back this conclusion. We there-
fore conclude that there are other study characteristics than the data structure which are deci-
sive for estimated effect sizes.
Drawing on the specific issues analyzed in the extensions, we first can confirm that bilateral
tax rates indeed do not only raise primary tax coefficients but also the measured significance
of tax influences on FDI. In contrast, public spending still does not influence primarily identi-
fied t-values. It thus neither affects semi-elasticities nor their significance. From a theoretical
50
point of view this is somewhat disappointing. However, one might conclude that the search
for appropriate proxies of relevant public inputs should not yet be abandoned. Interestingly,
country fixed effects do generally not have a significant influence on primary t-values. Ob-
viously, tax burden measures in most studies show sufficient variation and are not absorbed
into more general time-invariant location characteristics. It thus seems indeed advisable to
control for location fixed effects especially in studies based on aggregate data, since this cor-
rects some of the bias in measured tax coefficients (see again Table 12) without sacrificing
significance of the tax effect. Time fixed effects significantly reduce the significance of pri-
mary tax effects. Tax effects are thus rather absorbed by temporary macroeconomic condi-
tions within countries or regions than by static location characteristics. All these effects are
robust to the extension towards the full model. Moreover, we find that the coefficients for the
way of finance and openness of the target economy are fully in line with the prior results for
semi-elasticities, i.e. their effects on primary tax coefficients and significance are congruent.
Some of the remaining regressors (e.g. manufacturing industry data) seem to exert significant
effects on semi-elasticities, but not on t-values or the other way around.
As concerns the assessment of publication bias we have to rely on FEVD estimates performed
as a general robustness check in Column (3) of Table 13. Obviously, unpublished studies pro-
duce significantly lower t-values than published work. So this might not be in line with the
conclusions we drew from Column (2) of Table 12. But note that current publication status
might only reflect a temporary stage of the publication process and most researchers would
probably not want to adjust results before or during the following review process. So the in-
fluence of publication status might rather reflect other effects adherent to relatively new work
which was not adequately captured by the study-specific effects of our meta-regression than
different (implicit) requirements as concerns significance for published and unpublished re-
search. Therefore more weight should be given to the findings in Column (2) of Table 12 as
this is clearly the more established way to cope with the issue. As a result, publication bias is
not yet considered as unambiguously verified.
Strikingly, the influence of geographic FDI target regions (“Europe only” and “US only”) on
effect size estimates which we identified on the basis of the regressions based on semi-elasti-
cities is confirmed only for the United States. Investors targeting this economy apparently do
not base their decisions on tax aspects.
Finally, the results from FGLS estimation in Column (5) of Table 13 should be addressed.
Most results are in line with those obtained from the fixed effects cluster econometric estima-
51
tion. There are, however, also some differences. While the coefficient for country fixed ef-
fects is positive and significant, FGLS identifies no significant effect for time fixed effects.
Also the results for macroeconomic average tax rates are of opposite sign and significant. Fur-
thermore, no significant impact on primary t-values is identified for the bilateral average tax
rate (p = 0.106). But remember that the evidence against simple pooled estimation from our
series of tests is fairly pronounced. Thus, the pooled FGLS results are put into perspective. In
addition, choosing the optimal meta-regression estimator is not a minor issue with respect to
results, but represents a precondition for valid and reliable inference based on meta-data.
8. Conclusions
Tax competition and tax harmonization still are among the central themes in public econom-
ics. Proof of this contention is the ever increasing amount of theoretical and empirical studies.
Meanwhile, meta-analyses of these empirical studies exist. This paper does not simply present
another meta-analysis. In addition to an extension of the data base and a deeper methodologi-
cal discussion and approach, the main interest of our paper is the moderating influence of con-
trol variables, in particular public spending and agglomeration effects, on the estimated tax
rate effects on FDI as they could be obtained from the existing empirical studies.
According to our meta-analysis, a pooled effect based on medians, which seems rather robust
to different variations, could be obtained that amounts to a semi-elasticity for company taxes
on FDI of 1.68 in absolute terms. The most precise estimate of semi-elasticities is 1.39 in ab-
solute terms. Both effects are highly significantly different from zero in statistical terms.
While they obviously are also economically significant, this range of semi-elasticities is not
implausibly high. Taxes matter for location decisions and FDI.
The meta-regressions reported in this paper indicate that the heterogeneity in the semi-elasti-
cities and the t-statistics from the different empirical studies can be fairly well explained by
characteristics of the studies. We do not find overwhelming support for publication bias in the
meta-regressions with semi-elasticities as dependent variable, but some publication bias in the
meta-regressions with t- and z-statistics as dependent variable. The use of aggregate data
leads to higher semi-elasticities, but significantly reduces t-values and thus implies less pre-
cise estimates. Using aggregate data on affiliates significantly increases both. The inclusion of
country fixed effects might correct for some of the potential bias in aggregate studies, and it
does not even lead to lower significances. Apparently tax rates are sufficiently time-variant.
Interestingly we find that including time fixed effects reduces the significance of tax effects.
52
Regarding the choice of tax rates, unilateral effective average tax rates do not lead to signifi-
cantly higher tax effect size estimates or significances as compared to the statutory rate. Mar-
ginal effective tax rates even yield lower effects. This is probably not due to lower tax sensi-
tivity of marginal investment decisions, because we find evidence that discrete investment
choices are less sensitive. The low effect for EMTR might thus stem from problems of identi-
fying marginal choices in reality. Bilateral tax rates best capture tax incentives and yield both
significantly higher effect size estimates as well as higher significances. Finance and home
country double taxation relief do not influence investment decisions.
Regarding the control variables, it is most interesting that primary estimates are not signifi-
cantly affected by the inclusion of public spending. According to most estimates in the litera-
ture, the spending side does not moderate the tax rate effects. The results reported by Bénas-
sy-Quéré et al. (2007) do thus not find general support. This might be due to the fact that only
crude measures of public inputs are used in many studies. Similarly agglomeration effects do
not have any robust significant effect on estimated tax effects on FDI in the primary studies.
Certainly, more research is needed to find out whether the provision of public goods is empir-
ically related to tax rate effects or whether this is really a pure tax competition game.
References
Altshuler, R., H. Grubert and T.S. Newlon (2001), Has U.S. Investment Abroad Become More Sensi-tive to Tax Rates?, in: J.R. Hines (ed.), International Taxation and Multinational Activity, Uni-versity of Chicago Press, Chicago, 8 – 32.
Baldwin, R.E. and P. Krugman (2004), Agglomeration, Integration and Tax Harmonization, European Economic Review 48, 1 – 23.
Bartik, T.J. (1985), Business Location Decisions in the United States: Estimation of the Effects of Unionization, Taxes, and Other Characteristics of States, Journal of Business & Economic Sta-tistics 3, 14 – 22.
Bartik, T.J. (1991), Who Benefits from State and Local Economic Development Policies?, W.E. Up-john Institute for Employment Research, Kalamazoo.
Becker, J., C. Fuest and T. Hemmelgarn (2006), How Does FDI React to Corporate Taxation? A Comment on Studies Using Aggregate Data, Working Paper, University of Cologne.
Beckmann, R. (2007), Profitability of Western European Banking Systems: Panel Evidence on Struc-tural and Cyclical Determinants, Deutsche Bundesbank Discussion Paper Series 2: Banking and Financial Studies, DP No. 17/2007.
Bellak, C. and M. Leibrecht (2007), Some Further Evidence on the Role of Effective Corporate In-come Taxes as Determinants of Foreign Direct Investment in Central and East European Coun-tries, in: National Tax Association (ed.), Proceedings of the National Tax Association Confe-rence 2006, Boston, 311 – 343.
Bellak, C., M. Leibrecht and J.P. Damijan (2007), Infrastructure Endowment and Corporate Income Taxes as Determinants of Foreign Direct Investment in Central and Eastern European Countries, LICOS Discussion Paper Series No. 193/2007.
53
Bénassy-Quéré, A., L. Fontagné and A. Lahrèche-Révil (2005), How Does FDI React to Corporate Taxation? International Tax and Public Finance 12, 583 – 603.
Bénassy-Quéré, A., N. Gobalraja and A. Trannoy (2007), Tax and Public Input Competition, Econom-ic Policy 19, 385 – 430.
Bera, A.K. and W. Sosa-Escudero (2008), Tests for Unbalanced Error-Components Models under Local Misspecification, Stata Journal 8, 67 – 78.
Bijmolt, T.H.A. and R.G.M. Pieters (2001), Meta-Analysis in Marketing When Studies Contain Mul-tiple Measurements, Marketing Letters 12, 157 – 169.
Billington, N. (1999), The Location of Foreign Direct Investment: An Empirical Analysis, Applied Economics 31, 65 – 76.
Bloningen, B.A. (2005), A Review on the Empirical Literature on FDI Determinants, NBER Working Paper No. 11299.
Bobonis, G.J and H.J. Shatz (2007), Agglomeration, Adjustment, and State Policies in the Location of Foreign Direct Investment in the United States, Review of Economics and Statistics 89, 30 – 43.
Bom, P.R.D. and J.E. Ligthart (2008), How Productive is Public Capital? A Meta-Analysis, CESifo Working Paper Series WP No. 2206 and CenTER Discussion Paper Series DP No. 2008-10, Revised Version March 2008.
Boskin, M.J. and G. Gale (1987), New Results on the Effects of Tax Policy on the International Loca-tion of Investment, in: M. Feldstein (ed.), The Effects of Taxation on Capital Accumulation, University of Chicago Press, Chicago, 201 – 219.
Broekman, P. and W.N. van Vliet (2001), Winstbelasting en Kapitaalstromen in de EU, Openbare Uitgaven 33, 46 – 53.
Brons, M.R.E. (2006), Meta-Analytical Studies in Transport Economics: Methodology and Applica-tions, Dissertation, Vrije Universiteit Amsterdam.
Brons, M., P. Nijkamp, E. Pels and P. Rietveld (2008), A Meta-Analysis of the Price Elasticity of Gasoline Demand. A SUR Approach, Energy Economics 30, 2105 – 2122.
Brueckner, J.K. (1979), Property Values, Local Public Expenditure and Economic Efficiency, Journal of Public Economics 11, 223 – 245.
Brueckner, J.K. (1982), A Test for Allocative Efficiency in the Local Public Sector, Journal of Public Economics 19, 311 - 331.
Brülhart, M., M. Jametti and K. Schmidheiny (2007), Do Agglomeration Economies Reduce the Sen-sitivity of Firm Location to Tax Differentials?, mimeo, Université de Lausanne.
Buettner, T. (2002), The Impact of Taxes and Public Spending on the Location of FDI: Evidence from FDI-Flows within Europe, ZEW Discussion Paper No. 02-17.
Buettner, T. and M. Ruf (2007), Tax Incentives and the Location of FDI: Evidence from a Panel of German Multinationals, International Tax and Public Finance 14, 151 – 164.
Buettner, T. and G. Wamser (2008), The Impact of Nonprofit Taxes on Foreign Direct Investment: Evidence from German Multinationals, forthcoming in: International Tax and Public Finance.
Cassou, S.P. (1997), The Link Between Tax Rates and Foreign Direct Investment, Applied Economics 29, 1295 – 1301.
Demekas, D.G., B. Horváth, E. Ribakova and Y. Wu (2007), Foreign Direct Investment in European Transition Economies - the Role of Policies, Journal of Comparative Economics 35, 369 – 386.
De Mooij, R.A. and S. Ederveen (2003), Taxation and Foreign Direct Investment: A Synthesis of Em-pirical Research, International Tax and Public Finance 10, 673 – 693.
De Mooij, R.A. and S. Ederveen (2005), Explaining the Variation in Empirical Estimates of Tax Elas-
54
ticities of Foreign Direct Investment, Tinbergen Institute Discussion Paper, No. 2005-108/3. De Mooij, R.A. and S. Ederveen (2006), What a Difference Does it Make? Understanding the Empiri-
cal Literature on Taxation and International Capital Flows, European Commission Economic Paper No. 261.
Desai, M.A., C.F. Foley and J.R. Hines (2004), Foreign Direct Investment in a World of Multiple Taxes, Journal of Public Economics 88, 2727 – 2744.
Devereux, M.P. (2006), The Impact of Taxation on the Location of Capital, Firms and Profit: A Sur-vey of Empirical Evidence, Working Paper from Oxford University Centre for Business Taxa-tion No. 702.
Devereux, M.P. and H. Freeman (1995), The Impact of Tax on Foreign Direct Investment: Empirical Evidence and the Implications for Tax Integration Schemes, International Tax and Public Finance 2, 85 – 106.
Devereux, M.P. and R. Griffith (1998), Taxes and the Location of Production: Evidence from a Panel of US Multinationals, Journal of Public Economics 68, 335 – 367.
Devereux, M.P. and R. Griffith (1999), The Taxation of Discrete Investment Choices, IFS Working Paper No. W98/16.
Devereux, M.P. and B. Lockwood (2006), Taxes and the Size of the Foreign-Owned Capital Stock: Which Tax Rate Matters?, Paper presented at the European Tax Policy Forum 2006, London.
Dunning, J.H. (1981), International Production and the Multinational Enterprise, Allen & Unwin, London.
Egger, P., S. Loretz, M. Pfaffermayr and H. Winner (2008), Bilateral Effective Tax Rates and Foreign Direct Investment, Working Paper from Oxford University Centre for Business Taxation No. 0802.
Erkel-Rousse, H. and D. Mirza (2002), Import Price Elasticities: Reconsidering the Evidence, Cana-dian Journal of Economics 35, 282 – 306.
Feld, L.P. (2000), Steuerwettbewerb und seine Auswirkungen auf Allokation und Distribution: Ein Überblick und eine empirische Analyse für die Schweiz, Mohr Siebeck, Tübingen.
Feld, L.P., T. Baskaran and J. Schnellenbach (2007), Fiscal Federalism, Decentralization and Econom-ic Growth: A Meta-Analysis, Unpublished Manuscript, University of Heidelberg.
Fuest, C., B. Huber and J. Mintz (2005), Capital Mobility and Tax Competition, Foundations and Trends in Microeconomics 1 (1), 1 – 62.
Gatti, D. (2008), Macroeconomic effects of ownership structure in OECD countries, Paris School of Economics Working Paper No. 2008 – 09.
Goldstein, H. (1995), Multilevel Statistical Models, 2nd ed., Edward Arnold, London. Gorter, J. and A. Parikh (2003), How Sensitive is FDI to Differences in Corporate Income Taxation
within the EU?, De Economist 151, 193 – 204. Grubert, H. and J. Mutti (1991), Taxes, Tariffs and Transfer Pricing in Multinational Corporate Deci-
sion Making, Review of Economics and Statistics 73, 285 – 293. Grubert, H. and J. Mutti (2000), Do Taxes Influence Where U.S. Corporations Invest?, National Tax
Journal 53, 825 – 837. Grubert, H. and J. Mutti (2004), Empirical Asymmetries in Foreign Direct Investment and Taxation,
Journal of International Economics 62, 337 – 358. Hajkova, D., G. Nicoletti, L. Vartia and K.-Y. Yoo (2006), Taxation, Business Environment and FDI
Location in OECD Countries, OECD Economics Department Working Paper No. 502. Hartman, D.G. (1984), Tax Policy and Foreign Direct Investment in the United States, National Tax
55
Journal 37, 475 – 488. Hedges, L.V. and I. Olkin (1985), Statistical Methods for Meta-Analysis, Academic Press, Orlando. Helpman, E. (1984), A Simple Theory of International Trade with Multinational Corporations, Journal
of Political Economy 92, 451 – 471. Helpman, E. (1985), Multinational Corporations and Trade Structure, Review of Economic Studies 52,
443 – 458. Hines, J.R. (1996), Altered States: Taxes and the Location of Foreign Direct Investment in America,
American Economic Review 86, 1076 – 1094. Hines, J.R. (1997), Tax Policy and the Activities of Multinational Corporations, in: A.J. Auerbach
(ed.), Fiscal Policy: Lessons from Economic Research, MIT Press, Cambridge, 401 – 445. Hines, J.R. (1999), Lessons from Behavioral Responses to International Taxation, National Tax Jour-
nal 52, 305 – 322. Hines, J.R. and E.M. Rice (1994), Fiscal Paradise: Foreign Tax Havens and American Business, Quar-
terly Journal of Economics 109, 149 –182. Horstmann, I.J. and J.R. Markusen (1992), Endogenous Market Structures in International Trade (na-
tura facit saltum), Journal of International Economics 32, 109 – 129. Hox, J. (2002), Multilevel Analysis. Techniques and Applications, Lawrence Erlbaum Associates,
Mahwah. Hunter, J.E. and F.L. Schmidt (1990), Methods of Meta-Analysis: Correcting Error and Bias in Re-
search Findings, Sage Publications, Newbury Park. Jacobs, O.H. and C. Spengel (1996), European Tax Analyzer, ZEW Wirtschaftsanalysen Bd. 11, No-
mos, Baden-Baden. Jarrell, S.B. and T.D. Stanley (1989), Meta-Regression Analysis: A Quantitative Method of Literature
Surveys, Journal of Economic Surveys 3, 54 – 67. Jarrell, S.B. and T.D. Stanley (1990), A Meta-Analysis of the Union-Nonunion Wage Gap, Industrial
and Labor Relations Review 44, 54 – 67. Jarrell, S.B. and T.D. Stanley (2004), Declining Bias and Gender Wage Discrimination? A Meta-Re-
gression Analysis, The Journal of Human Resources 39, 828 – 838. Jarrell, S.B. and T.D. Stanley (2005), Meta-Regression Analysis: A Quantitative Method of Literature
Surveys, Journal of Economic Surveys 19, 299 – 308. Jun, J. (1994), How Taxation Affects Foreign Direct Investment: Country Specific Evidence, World
Bank Policy Research Working Paper No.1307. Keen, M. and M. Marchand (1997), Fiscal Competition and the Pattern of Public Spending, Journal of
Public Economics 66, 33 – 53. King, M.A. and D. Fullerton (1984), The Taxation of Income from Capital, University of Chicago
Press, Chicago. Longhi, S., P. Nijkamp and J. Poot (2008), Meta-Analysis of Empirical Evidence on the Labour Mar-
ket Impacts of Immigration, IZA Discussion Paper Series No. 3418. Neumark, F. (1963), Tax Harmonization in the Common Market (‘Neumark-Report’), New York. Nichols, A. and M.E. Schaffer (2007), Clustered Standard Errors in Stata, Papers by Stata User Group
from United Kingdom Stata Users’ Group Meetings 2007, No. 07.
Markusen, J.R. (1997), Trade Versus Investment Liberalization, NBER Working Paper No. 6231. Markusen, J.R. (2002), Multinational Firms and the Theory of International Trade, MIT Press, Cam-
bridge.
56
Markusen, J.R., A. J. Venables, D. Eby-Konan and K. Honglin Zhang (1996), A Unified Treatment of Horizontal Direct Investment, Vertical Direct Investment and the Pattern of Trade in Goods and Services, NBER Working Paper No 5696.
Martens-Weiner, J. (2006), Company Tax Reform in the European Union: Guidance from the United States and Canada on Implementing Formulary Apportionment in the EU, Springer, New York.
McFadden, D. (1974), Conditional Logit Analysis of Qualitative Choice Behavior in: P. Zarembka (ed.), Frontiers in Econometrics, Academic Press, New York, 105 – 142.
McFadden, D. (1978), Modeling the Choice of Residential Location, in: A. Karlquist et al. (eds.), Spa-tial Interaction Theory and Planning Models, North-Holland, Amsterdam, 75 – 96.
Murthy, N.R.V. (1989), The Effects of Taxes and Rates of Return on Foreign Direct Investment in the United States: Some Econometric Comments, National Tax Journal 42, 205 – 207.
Newlon, T.S. (1987), Tax Policy and Multinational Firm’s Financial Policy and Investment Deci-sions, Ph. D. Thesis, Princeton University.
Overesch, M. and G. Wamser (2008a), The Effects of Company Taxation in EU Accession Countries on German Multinationals, Paper prepared for the Meeting of the European Tax Policy Forum, London.
Overesch, M. and G. Wamser (2008b), Who Cares about Corporate Taxation? Asymmetric Tax Ef-fects on Outbound FDI, ifo Working Paper No. 59.
Pain, N. and G. Young (1996), Tax Competition and the Pattern of European Foreign Direct Invest-ment, mimeo, National Institute of Economic and Social Research.
Papke, L.E. (1991), Interstate Business Tax Differentials and New Firm Locations: Evidence from Panel Data, Journal of Public Economics 45, 47 – 68.
Pluemper, T. and V.E. Troeger, Efficient Estimation of Time-Invariant and Rarely Changing Variables in Finite Sample Panel Analyses with Unit Fixed Effects, Policy Analysis 15, 124 – 139.
Razin, A., Y. Rubinstein and E. Sadka (2005), Corporate Taxation and Bilateral FDI with Threshold Barriers, NBER Working Paper No. 11196.
Schanz, G. von (1890), Die Steuern der Schweiz in ihrer Entwicklung seit Beginn des 19. Jahrhun-derts, Vol I to V, Cotta’sche Buchhandlung, Stuttgart.
Schmenner, R.W. (1982), Making Business Location Decisions, Prentice-Hall Inc., Englewood Cliffs. Schön, W., U. Schreiber and C. Spengel (eds.) (2008), A Common Consolidated Corporate Tax Base
for Europe, Springer, Berlin and Heidelberg. Shang-Jin, W. (1997), How Taxing is Corruption on International Investors? NBER Working Paper
No. 6030.
Sinn, H.-W. (2003), The New Systems Competition, Blackwell, Oxford. Slemrod, J. (1990), Tax Effects on Foreign Direct Investment in the U.S.: Evidence from a Cross-
Country Comparison, in: A. Razin and J. Slemrod (eds.), Taxation in the Global Economy, Uni-versity of Chicago Press, Chicago, 79 – 122.
Spengel, C. (1995), Europäische Steuerbelastungsvergleiche: Deutschland – Frankreich – Großbri-tannien, IDW-Verlag, Düsseldorf.
Spoerer, M. (2002), Wann begannen Fiskal- und Steuerwettbewerb? Eine Spurensuche in Preußen, anderen deutschen Staaten und der Schweiz, Jahrbuch für Wirtschaftsgeschichte 2, 11 – 35.
Stanley, T.D. (1998), New Wine in Old Bottles: A Meta-Analysis of Ricardian Equivalence, Southern Economic Journal 64, 713 – 727.
Stanley, T.D. (2000a), Challenging Time Series: Limits to Knowledge, Inertia and Caprice, Edward Elgar, Cheltenham.
57
Stanley, T.D. (2000b), An Empirical Critique of the Lucas Critique, Journal of Socio-Economics 29, 91 – 107.
Stanley, T.D. (2001), Wheat from Chaff: Meta-Analysis as Quantitative Literature Review, Journal of Economic Perspectives 15, 131 – 150.
Stanley, T.D. (2005), Beyond Publication Bias, Journal of Economic Surveys 19, 309 – 345. Stanley, T.D. (2008), Meta-Regression Methods for Detecting and Estimating Empirical Effects in the
Presence of Publication Selection, Oxford Bulletin of Economics and Statistics 70, 103 – 127. Stanley, T.D. and S.B. Jarrell (1989), Meta-Regression Analysis: A Quantitative Method of Literature
Surveys, Journal of Economic Surveys 3, 54 – 67. Stanley, T.D. and S.B. Jarrell (1998), Gender Wage Discrimination Bias? A Meta-Regression Analy-
sis, Journal of Human Resources 33, 947 – 973. Stöwhase, S. (2005a), Tax-Rate-Differentials and Sector-Specific Foreign Direct Investment: Empiri-
cal Evidence from the EU, FinanzArchiv 61, 535 – 558. Stöwhase, S. (2005b), Taxes and Multinational Enterprises in the EU: Location Decisions and Income
Shifting, Dissertation, LMU München. Swenson, D.L. (1994), The Impact of U.S. Tax Reform on Foreign Direct Investment on Foreign Di-
rect Investment in the United States, Journal of Public Economics 54, 243 – 266. Swenson, D.L. (2001), Transaction Type and the Effect of Taxes on the Distribution of Foreign Direct
Investment in the United States, in: J.R. Hines (ed.), International Taxation and Multinational Activity, University of Chicago Press, Chicago, 89 – 112.
Thompson, S.G. and S.J. Sharp (1999), Explaining Heterogeneity in Meta-Analysis: A Comparison of Methods, Statistics in Medicine 18, 2693 – 2708.
Tiebout, C.M. (1956), A Pure Theory of Local Expenditures, Journal of Political Economy 64, 416 – 424.
Volckart, O. (2002), Wettbewerb und Wettbewerbsbeschränkung im vormodernen Deutschland 1000 – 1800, Mohr Siebeck, Tübingen.
Wijeweera, A., B. Dollery and D. Clark (2007), Corporate Tax Rates and Foreign Direct Investment in the United States, Applied Economics 39, 109 – 117.
Wilson, J.D. (1999), Theories of Tax Competition, National Tax Journal 52, 269 – 304. Wilson, J.D. and D.E. Wildasin (2004), Capital Tax Competition: Bane or Boon?, Journal of Public
Economics 88, 1065 – 1091. Wolff, G.B. (2007), Foreign Direct Investment in the Enlarged EU: Do Taxes Matter and to What
Extent?, Open Economies Review 18, 327 – 346. Wooldridge, J.M. (2002), Econometric Analysis of Cross Section and Panel Data, MIT Press, Cam-
bridge. Wooldridge, J.M. (2003), Cluster-Sample Methods in Applied Econometrics, American Economic
Review 93, 133 – 138. Young, K.H. (1988), The Effects of Taxes and Rates of Return on Foreign Direct Investment in the
United States, National Tax Journal 41, 109 – 121. Zodrow, G.R. and P. Mieszkowski (1986), Pigou, Tiebout, Property Taxation and the Underprovision
of Local Public Goods, Journal of Urban Economics 19, 356 – 370.